The browser, native app or messaging client you actually use from the phone.
A replaceable stack, not a shopping list
Which AI Tool Fits Your Workflow and Data Boundary?
Search practical local, hosted and agent tools by what they do, where they run, how they are licensed and the caveat most likely to change your decision—including whether a phone can be the controller.
179searchable entriesPRIMARY LINKS + CAVEATS
22practical topicsJOB BEFORE BRAND
0automatic endorsementsVERIFY THE EXACT VERSION
Vetted tool catalogue
How do you choose the right AI tool?
These are replaceable examples, not endorsements or a required stack. Each entry names the practical fit and the caveat that most changes the decision.
Authenticates the person and device, owns sessions, constrains tools and routes requests.
Loads the model or provides retrieval and automation. It is not a mobile interface by itself.
Keep phone, home and cloud visible separately; “local model” does not describe the transport.
Search names, jobs, platforms, licenses and caveats. Search combines with the topic filter below.
Filter by topic
Run models locally
KTransformers
An Apache-2.0 research framework for running and fine-tuning large mixture-of-experts models across CPU and GPU memory. Its current path exposes hardware-specific kernels and LLaMA-Factory integration; follow the exact model and instruction-set tutorial, and treat project throughput figures as configuration-specific rather than a desktop promise.
CPU + GPUMoE modelsResearch project
Ollama + Open WebUI
A straightforward local API and browser interface for trying multiple models. Mobile role: open the authenticated WebUI from the phone over a private route while Ollama and the model stay on the host. Audit exposed network interfaces, user access and any cloud fallback. Current Open WebUI releases use a branding-restricted license that is not OSI-approved; call it self-hostable, not strict OSS.
Phone browserHome inferenceLicense caveat
LibreChat
A multi-user web interface for local and hosted models, custom OpenAI-compatible endpoints, agents, files and MCP tools. It fits in front of vLLM or several providers; treat authentication, MCP credentials and code execution as production services rather than desktop conveniences.
Self-hostedMulti-providerAgents + MCP
SillyTavern
An AGPL-3.0, locally installed interface for switching among local and hosted text models, image generators and speech engines, with detailed prompt, character, lorebook and extension controls. It is deliberately flexible rather than beginner-simple. Keep it on a trusted network: the project warns against exposing the server directly, stores multi-user data and API keys in plain text on the server, and gives UI extensions broad access while server plugins are unsandboxed.
Local interfaceMany backendsExtensions need trust
llama.cpp
A portable, low-level local inference foundation with broad quantization and hardware support. Best when you want control and can own the configuration.
LocalTechnicalPortable
Project repository →LM Studio
A desktop route for downloading, comparing and serving local models without assembling a full command-line stack. Mobile role: backend only—enable API-token authentication, keep the server private and put a phone-friendly interface or gateway in front.
Home backendAPI token availablePrivate route
vLLM
A high-throughput, OpenAI-compatible serving engine for shared GPU inference. Mobile role: remote-ready backend, not the public phone endpoint; keep it behind a private network and authenticated gateway. Use it when concurrency and batching matter; for DGX Spark follow NVIDIA's ARM64/Blackwell recipe and current model matrix instead of generic x86 installation instructions.
Home/server backendOpenAI-compatible APIGateway required
SGLang
An Apache-2.0 serving framework for language and multimodal models, from one GPU to distributed clusters. Its runtime covers prefix caching, continuous batching, prefill-decode disaggregation, speculative decoding, structured outputs and multiple parallelism strategies, with integrations for reinforcement-learning rollout generation. Treat project performance claims as configuration-specific: pin the model, engine, kernels, hardware and traffic shape, then measure latency, throughput, correctness and recovery on your workload.
GPU serviceDistributedRL integration
TokenSpeed
An MIT-licensed, OpenAI-compatible inference engine from the LightSeek Foundation, built around a compiler-backed model layer, C++ scheduler and pluggable kernels for agentic and large-MoE workloads. Current recipes cover Kimi K3, DeepSeek V4, GLM‑5.2, GPT‑OSS and other model families across selected NVIDIA and AMD systems. This is an operator-grade, fast-moving stack—not a generic desktop runtime: pin the checkpoint, container, kernel bundle and topology, then validate correctness, context capacity and throughput together.
GPU serviceNVIDIA + AMD recipesFast-moving
Superlinked SIE
An Apache-2.0 self-hosted server and Kubernetes stack that puts embedding, reranking, document conversion, structured output, content screening and text generation behind one OpenAI-compatible API. It loads models on demand through task-specific CPU and NVIDIA bundles, with a local Apple-silicon route. Pin the model weights and container or chart version, validate each model’s licence and hardware fit, and account for first-call downloads and opt-out usage telemetry before production.
Apache-2.0Local + KubernetesMany model tasks
LiteLLM Proxy
A central OpenAI-compatible gateway for multiple local and hosted endpoints, with authentication hooks, budgets, rate limits and spend tracking. Mobile role: it can keep one endpoint stable while a phone session routes between home and cloud models, but a public app still needs end-user identity and session handling in front. It becomes a security boundary: keep it patched, authenticated and private.
Hybrid routingBudgetsShared service
Official documentation →Mistral
Use Vibe for an end-user assistant, Studio/API for applications, or supported open-weight models through a local runtime. Availability and licenses differ by model; “Mistral” is not one deployment or one data boundary.
Hosted + localOpen-weight optionsEU company
Compare Mistral routes →Foundry Local
Microsoft's GA runtime embeds curated ONNX Runtime-optimized chat and speech models into Windows, Apple-silicon macOS and Linux applications. C#, JavaScript, Python and Rust SDKs manage model download, hardware-specific variants and in-process inference; an optional OpenAI-compatible server supports local integrations. It needs no Azure subscription or API key, and it is designed for single-user on-device apps—not hosted Microsoft Foundry or multi-user serving like vLLM. The SDK is MIT-licensed, the CLI uses Microsoft terms and each model keeps its own license.
On-deviceC# + JS + Python + RustCurated models
Open WebUI Pipelines
Run Python filters, provider adapters and compute-heavy pre/post-processing behind an OpenAI-compatible endpoint. The project itself recommends built-in Functions for simple filters. Pipelines execute arbitrary code and are neither a durable scheduler nor a permission boundary.
PythonFilters + adaptersNot orchestration
Pipelines repository →LLM CLI
Simon Willison's Apache-2.0 command-line tool and Python library can call hosted, OpenAI-compatible and locally installed models, extract structured data, create embeddings and invoke tools through plugins. It logs prompts and responses to a local SQLite database by default; turn logging off or set retention and access controls before processing sensitive material. Plugins are executable Python packages, so pin and review them rather than treating the directory as a trust list.
Apache-2.0Hosted + local modelsLocal history by default
Operate AI infrastructure
NVIDIA NIM + NeMo Framework + RAPIDS
NIM packages supported model-serving APIs, NeMo supplies model-development and customization workflows, and RAPIDS accelerates GPU data science. They are complementary products rather than one required bundle; confirm the supported model, container, hardware and commercial entitlement for the exact deployment.
Model APIsDevelopmentGPU analytics
TensorRT-LLM + Triton Inference Server + NVIDIA Dynamo
TensorRT-LLM optimizes supported large-model inference, Triton exposes production model-server endpoints and Dynamo coordinates distributed inference. Select only the components the service needs, then pin the model, kernels, driver, container and topology used in acceptance testing.
OptimizationServingDistributed inference
NVIDIA Run:ai + Mission Control + Base Command Manager
Run:ai schedules and governs shared GPU workloads, Mission Control manages supported AI-factory configurations, and Base Command Manager provisions and operates clusters. Licensing, platform support and control-plane requirements differ; verify the exact product release and order rather than assuming hardware includes them.
SchedulingFleet operationsEntitlement-sensitive
NVIDIA Omniverse + AI Workbench
As of May 2026, Omniverse is free for development, production and redistribution without NVIDIA AI Enterprise; free use has community support, while enterprise support needs the applicable subscription. NVIDIA AI Workbench is also free for development workflows. Neither is a free-forever entitlement to the complete supported NVIDIA AI Enterprise production suite.
DevelopmentCommunity supportNot the full suite
NVIDIA UFM + NetQ
UFM manages supported InfiniBand fabrics while NetQ observes and troubleshoots supported Ethernet and network environments. They cover different fabrics and support matrices; confirm topology, device software, licensing and telemetry retention before choosing either.
InfiniBandNetwork telemetrySupport matrix
DGX OS + DCGM + CUDA + NVIDIA Container Toolkit
DGX OS supplies the supported system baseline, DCGM exposes GPU health and telemetry, CUDA provides the GPU programming platform, and Container Toolkit gives containers controlled GPU access. Pin their compatibility chain; installing these components does not by itself prove an NVIDIA AI Enterprise entitlement.
System softwareGPU telemetryContainers
Kubernetes + Kueue + Volcano + Slurm
Kubernetes is the container-orchestration baseline, Kueue adds quota-aware batch admission, Volcano adds Kubernetes batch scheduling, and Slurm supplies an established HPC workload manager. Choose one primary queue and scheduling authority; overlapping control loops make priority, recovery and capacity ownership harder to reason about.
GPU schedulingBatch queuesOperator-owned
PyTorch + Transformers + PEFT + TRL + MLflow
PyTorch supplies the training framework; Transformers provides model implementations; PEFT supplies parameter-efficient adapters; TRL provides supervised and preference-training loops; and MLflow records runs and artifacts. Pin the model, dataset, evaluation and environment together: this is a composable lifecycle stack, not a pre-integrated production platform.
TrainingAdapters + preferenceExperiment evidence
NVIDIA GPU Operator + KAI Scheduler
NVIDIA GPU Operator manages the Kubernetes software components needed to run NVIDIA GPUs; KAI Scheduler adds AI-oriented scheduling, including gang scheduling, to the workload-control layer. They do not replace a clear queue owner, compatibility testing or the hardware-vendor support boundary.
Kubernetes GPUsGang schedulingOperator-owned
Argo Workflows + Harbor
Argo Workflows runs Kubernetes-native DAG and step workflows; Harbor provides a private artifact registry. Pair them only with explicit access, signing, retention and promotion rules—an image in a registry is not a safe or approved deployment by itself.
Workflow automationPrivate registryPolicy required
MAAS + Foreman + Ansible
MAAS and Foreman are alternative routes for provisioning and host lifecycle management; Ansible applies repeatable configuration and operational automation. None removes the need for tested backups, firmware coordination, break-glass access and a documented rebuild path.
Bare metalConfigurationRecovery ownership
Ceph + MinIO + Cilium + Vault
Ceph supplies distributed object, block and file storage; MinIO's current vendor route is AIStor while the former community repository is archived; Cilium provides eBPF networking and policy; and Vault manages secrets under HashiCorp's current terms. Treat this as four separately operated services, not a pre-integrated platform.
StorageNetwork policySecrets
Redfish + NVIDIA Driver + NCCL + platform firmware
Redfish standardizes server-management APIs, the NVIDIA driver connects the operating system to the accelerator, NCCL coordinates collective communication and platform firmware remains OEM-specific. CUDA is catalogued with the NVIDIA system foundation above. Validate the whole compatibility chain and support route before an upgrade.
Hardware APIGPU runtimeOEM boundary
Choose model and embedding services
Cohere + Cohere Embed
Cohere provides hosted chat, embedding and reranking APIs. Treat each endpoint as a separate design choice: record the model version, input type and data boundary, and evaluate retrieval with your own corpus before committing to an index.
Hosted APIEmbed + rerankVersion explicitly
Nomic Embed + Sentence Transformers
Nomic Embed v2 is an Apache-2.0 multilingual embedding model that runs locally through Sentence Transformers or Transformers. Benchmark retrieval on your own languages, use the required query and document prefixes, and pin the model revision and output dimension—changing either means re-embedding the corpus.
Local weightsMultilingual retrievalPin the index contract
Models.dev
An MIT-licensed, community-maintained catalogue and JSON API for model identifiers, provider routes, context limits, capabilities and prices, with a type-safe SDK and offline snapshot. Use it to seed selectors, routing tests and cost comparisons—not as the final procurement source. Pin a snapshot and reconcile consequential limits, regional availability, terms and prices with each provider's current documentation.
MITJSON API + SDKVerify with providers
Build speech and calling
Microsoft VibeVoice
An MIT-licensed codebase for long-form speech recognition and text to speech. The project publishes an ASR model with speaker, timestamp and transcript structure plus a real-time TTS model; model weights keep their own terms. Test language, diarization and timestamp accuracy on your audio, and require consent for cloned or synthetic voices.
ASR + TTSLocal modelsConsent required
VibeVoice repository →Voicebox
An MIT-licensed desktop stack for voice cloning, speech generation, dictation and MCP voice output across several model engines. It can keep captures and models local, but GPU support, weight licenses and quality vary by engine. Clone only voices you own or have permission to use, and keep disclosure and anti-impersonation controls around generated audio.
Local-firstSTT + TTSVoice cloning
OpenFlow
A free MIT-licensed voice-to-text app for macOS and Windows that can type into the active application. Speech recognition and optional AI cleanup can each run locally, through a self-hosted endpoint or through a remote provider, so confirm both settings before calling a workflow private. Current downloads are unsigned and trigger an operating-system warning; verify the release or build from source rather than treating the bypass as routine.
Local-firstmacOS + WindowsUnsigned builds
Superwhisper
A proprietary dictation app for macOS, Windows and iOS with custom vocabulary, task-specific modes, file transcription and meeting notes. It offers local and cloud speech and language models and can work offline, but privacy and connectivity depend on the models and mode you select; the vendor notes that offline models perform best on Apple-silicon Macs.
Local + cloudDesktop + iOSFree + paid tiers
Asterisk + FreeSWITCH
Connect a local operator's SIP trunk to an open-source PBX, then control outbound calls, DTMF, media and hangup events through Asterisk ARI or FreeSWITCH ESL. The carrier, number rights, emergency restrictions and call rates remain contractual services.
Self-operatedSIP/RTPCall control
Speaches
An MIT-licensed, OpenAI-compatible server for streaming transcription, translation and speech generation. It combines faster-whisper for STT with Kokoro or Piper-family TTS paths and supports CPU or GPU deployment.
MITOpenAI-compatibleStreaming
Kokoro-FastAPI + openedai-speech
Kokoro-FastAPI is the active Apache-licensed, OpenAI-compatible TTS wrapper to test first. openedai-speech remains a useful AGPL reference for Piper/XTTS and voice mapping, but its maintainer archived it in January 2026 and calls it mostly obsolete.
Local voicesOpenAI speech APIArchive noted
Pipecat + LiveKit Agents
Open-source frameworks for streaming transports, VAD, interruptions, STT, LLM and TTS stages. Use them for the media conversation loop; keep appointment locks and business state in PostgreSQL and a durable workflow engine.
Real-time mediaPluggable speechBarge-in
Work with documents and media
DeepSeek-OCR
A research release for turning document images into compact textual context through vLLM or Transformers. The repository is MIT-licensed and its reference environment pins CUDA 11.8, PyTorch 2.6 and FlashAttention. Treat “optical compression” as a model technique—not lossless storage—and measure transcription, reading order, tables and hallucinations against page images.
OCRGPU-orientedVerify against source
OfficeCLI
An Apache-2.0 command-line tool for agents to read and edit Word, Excel and PowerPoint files without an Office installation. It offers a local preview and single-binary releases. Work on copies, render the result and compare formulas, charts, fonts and pagination in a real Office-compatible viewer; ZIP-format access does not guarantee perfect application fidelity.
DOCX + XLSX + PPTXNo Office installRender and verify
OfficeCLI repository →ComfyUI + IOPaint
ComfyUI is a GPL-3.0 node-graph engine for repeatable local image, video, audio and 3D workflows; IOPaint is an Apache-2.0, self-hosted route for erasing, replacing and extending images on CPU, GPU or Apple silicon. Model weights and outputs keep their own licences. Pin stable releases and known workflows, keep ComfyUI's optional API nodes disabled for a fully offline path, and treat third-party models and custom nodes as untrusted code and artifacts.
Local + self-hostedVisual workflowsAudit nodes + weights
FramePack Studio
An Apache-2.0 local application based on FramePack for image-to-video and text-to-video work, with queueing, prompt timelines and blending, LoRA, upscale and post-processing paths. It requires a CUDA-compatible GPU: 8 GB VRAM is the stated minimum, while 16 GB or more and 80 GB or more of storage are recommended.
Local videoCUDA GPU8 GB VRAM minimum
Amuse
A local Windows studio for image, video, audio and text pipelines plus editing, upscale and interpolation. Its current reference route uses CUDA 13 and recommends RTX hardware. The software licence is personal and non-commercial; commercial use needs a separate licence, and model weights keep their own terms.
WindowsCUDA 13 + RTXCommercial licence needed
Pinokio
An MIT-licensed local launcher and runtime for installing and running AI apps. Its scripts can execute commands and download code with your authority: review source, prefer verified or frozen scripts, and isolate credentials before running an app.
Local launcherMITScripts need review
threestudio
An Apache-2.0 local research framework for text, image and few-shot 3D generation across multiple methods and extensions. It is NVIDIA/CUDA-oriented, Ubuntu-tested and states a 6 GB VRAM minimum. Dependencies and model licences vary, and this is a research stack rather than a beginner desktop default.
NVIDIA + CUDA6 GB VRAM minimumResearch stack
Build AI applications and agents
OpenManus
An MIT-licensed implementation of a general tool-using agent with browser, code, retrieval, sandbox and multi-agent paths. The maintainers describe it as a simple prototype. Model endpoints, browser access and executable tools create the real data and permission boundary, so start in a disposable environment with narrow credentials and an external acceptance test.
PythonGeneral agentPrototype
OpenManus repository →CrewAI + CAMEL
CrewAI packages agents, tasks, crews and stateful flows; CAMEL focuses on multi-agent research and composable agent societies. Use either only when separate roles improve a measured result—more agents add coordination cost, failure modes and permissions.
PythonMulti-agentRole-based
TinyTroupe
Microsoft's MIT-licensed Python library simulates configurable personas and focus groups for research and product exploration. It runs on your machine, but OpenAI and Azure OpenAI remain the primary model routes and Ollama support is experimental. Treat synthetic responses as hypotheses, validate them against real people and data, and read the project's legal disclaimer before use.
Local libraryPersona simulationResearch only
TinyTroupe repository →LlamaIndex Workflows + Haystack
LlamaIndex Workflows provides event-driven, step-based control for retrieval and agents. Haystack composes model, retrieval and processing components into explicit pipelines. Both suit knowledge-heavy applications; keep source provenance and evaluation independent of the framework. Plan the retrieval architecture.
PythonRetrievalExplicit pipelines
Agno + Pydantic AI
Agno supplies a broad agent stack with teams, workflows and knowledge integrations. Pydantic AI emphasizes typed dependencies, validated outputs, tools and testable Python application patterns. Choose the smaller abstraction that meets the acceptance test and keep business authorization outside the model loop.
PythonStructured outputTools + workflows
DSPy
Define model programs from typed modules and signatures, then optimize prompts or weights against examples and metrics. DSPy is most useful when you have a representative dataset and a meaningful objective; optimization against a weak metric simply automates overfitting.
Code-firstOptimizationEvaluation required
DSPy documentation →FastMCP + Official MCP Registry
FastMCP helps build MCP clients and servers; the official registry publishes discoverable server metadata. A registry listing is not a security review: pin packages, inspect requested capabilities, isolate credentials and test every tool's authorization boundary.
MCPPythonDiscovery, not trust
agent-browser
An Apache-2.0 native Rust CLI that automates a browser locally and can start a local MCP stdio server for an agent. It can inspect the accessibility tree, navigate, click, fill forms, upload files, and read cookies or network requests. The browser still opens real services under a real signed-in session: isolate the profile and credentials, allow only authorized domains and accounts, require confirmation for uploads or consequential actions, and do not mistake local execution for a local data boundary. First setup downloads Chrome for Testing unless a compatible browser is already available.
Apache-2.0CLI + local MCPLive-session authority
Use agent skills and playbooks
Agent Skills specification
A lightweight open format in which a folder contains a required SKILL.md plus optional scripts, references and assets. It provides a portable packaging convention, not a trust or permission boundary: read the instructions and executable files, pin the source and test the skill with least privilege before installing it broadly.
Open formatPortable instructionsInspect before use
Agent Skills overview →Microsoft SkillOpt
An MIT-licensed optimizer that revises natural-language skills from agent trajectories and promotes an artifact only through validation gates. It changes prompts and procedures rather than model weights. Keep an untouched holdout, version every candidate and include model/API cost and data exposure in the experiment; repeated optimization can overfit a weak evaluator.
PythonSkills + evalsHoldout required
SkillOpt repository →Claude Code Apple Skills
An MIT-licensed collection covering iOS, macOS, product work, testing, App Store preparation and code generation. Marketplace installs track main; use the repository's era tags when reproducibility matters. Treat generated patterns and legal copy as starting points, then verify them against the target Xcode, SDK, platform guidance and counsel where needed.
Apple platformsPrompt libraryPin for stability
Apple Skills repository →Superpowers
An MIT-licensed, cross-agent skill set that enforces discovery, specification, planning, test-driven implementation and review stages. It is an opinionated engineering method rather than evidence that autonomous coding is reliable. Adopt the parts that improve your acceptance tests and remove roles or ceremony that do not improve measured outcomes.
PlanningTDDCross-agent
Superpowers repository →Agency Agents
An MIT-licensed catalogue of role prompts spanning engineering, design, marketing, finance, healthcare and other functions. These are reusable briefs and personalities—not verified professionals or validated operating procedures. Test deliverables against primary sources and named reviewers, especially in regulated, financial, legal or health work.
Prompt libraryMany domainsNo credential guarantee
Agency Agents repository →Claude for Legal
Anthropic's Apache-2.0 reference agents, skills and connectors for commercial, privacy, product, corporate, employment, litigation, regulatory and learning workflows. They can run as Claude plugins or behind the Managed Agents API. They are not legal advice: preserve privilege, apply the correct jurisdiction, minimize connector access and require qualified counsel to release consequential work.
Claude ecosystemLegal workflowsHuman counsel
Claude for Legal repository →Claude SEO
An MIT-licensed Claude Code plugin with technical, content, schema, local, international and AI-search audit workflows. The repository says its recommendations are grounded in primary Google guidance, while optional providers add external data collection. Recheck every recommendation, respect robots and service terms, and treat rankings or traffic as measured outcomes—not promises.
SEO auditsClaude CodeOptional data APIs
Claude SEO repository →gstack
An MIT-licensed set of Claude Code skills for product, design, engineering, QA, security review and release work. Its productivity examples are the author's own measurements, not an independent benchmark. Some workflows can browse, change code or deploy, so inspect each skill, narrow credentials and keep repository review, tests and release approval outside the persona.
Claude CodeProduct to releaseSelf-reported results
gstack repository →Build knowledge and workflows
Hermes unbroker
An MIT-licensed Hermes skill that finds a consenting person's exposure on people-search sites, submits supported opt-outs and records human-only steps. It is US-first and may use browser sessions, email credentials and a dossier containing the exact personal data being removed. Encrypt state, minimize disclosure, verify consent and removals, and recheck legal and site requirements.
Privacy workflowHermes skillSensitive dossier
unbroker skill →OpenWiki
An MIT-licensed CLI that creates and updates codebase documentation or a personal local wiki in a portable Markdown format. The configured model can receive repository or note context, and scheduled runs send anonymous reliability telemetry unless disabled. Review generated changes like code: pin the provider and prompt, require diffs and keep primary documentation canonical.
Code docs + personal wikiMarkdownReview updates
OpenWiki repository →TaxHacker
An MIT-licensed application for extracting fields from receipts, invoices and transactions into a structured accounting view. Self-hosting the app does not make a remote model local, and the documents contain financial and personal data. Keep originals and your accounting system canonical, configure the model boundary explicitly and require human review for categories, tax and filings.
Receipts + invoicesSelf-hosted appNot accounting advice
TaxHacker repository →Firecrawl + Scrapy
Firecrawl packages search, scrape and crawl operations behind a self-hostable API; its core is AGPL-3.0 and the hosted service includes additional features. Scrapy is a BSD-licensed Python framework for explicit spiders and extraction pipelines. Whichever route you choose, respect robots directives, site terms, privacy and rate limits, and store source URLs and retrieval times with extracted content.
Self-hostedWeb crawlingPolicy required
Obsidian
A free-to-use, proprietary app that stores the canonical vault locally as plain-text Markdown. Backlinks, properties and graph view help navigate relationships; they do not make the vault a neural network. Treat community plugins as executable code and audit optional Sync, Publish and AI endpoints separately.
Local filesPortable MarkdownPlugin boundary
screenpipe
A source-available desktop recorder that indexes screen text, screenshots, app and browser context, audio transcripts and user input locally, then exposes search, an API, MCP and scheduled agent workflows. That breadth creates a high-risk monitoring dataset even when storage stays on-device. For workplace use, define purpose and legal basis, consult workers where required, complete a DPIA where systematic monitoring triggers it, minimize captured apps and fields, set short retention and deletion paths, encrypt storage, disable default analytics and keep cloud transcription, sync and hosted models off unless their processors and transfers are approved.
Local by defaultSource-availableDPIA likely at work
Chonkie
An MIT-licensed Python library for turning text, Markdown, tables and code into retrieval chunks. It offers token, sentence, recursive, semantic, neural and code-aware chunkers plus optional embedding and vector-store integrations. Preserve file and heading provenance, and evaluate retrieval on your documents instead of trusting a chunking benchmark. Follow the RAG build sequence.
MITLocal-capableReplaceable pipeline
FaultLine
An early-stage AGPL service that exposes correctable, structured facts through MCP and OpenWebUI tools. Its repository describes a validation gate, confidence strengthening and archived corrections; treat those as project claims to test. Keep source notes canonical, secure the MCP endpoint and verify user isolation, backup and retraction behavior.
AGPL-3.0Local serviceMCP fact graph
Docling + AnythingLLM
Parse complex files with provenance, then retrieve over a bounded collection with a model endpoint you can replace. See how the stages fit together.
Local-capableRAGPage provenance
RAGFlow + Embedchain
RAGFlow is a self-hostable RAG platform with document processing, retrieval and agent workflows; Embedchain is a lighter Python framework for adding data and querying an application. Use them to accelerate a prototype, but retain the original files, chunk metadata and an exportable evaluation set. Define the acceptance test.
RAGApplication layerKeep provenance
Unstructured + Microsoft GraphRAG
Unstructured partitions and stages documents for downstream retrieval. Microsoft GraphRAG builds graph-based indexes and query paths for relationship-heavy corpora. Neither removes the need to inspect extraction quality, preserve page-level provenance and compare the result with simpler retrieval. Compare RAG routes.
IngestionGraphRAGEvaluate complexity
Qdrant
Add Qdrant when retrieval needs explicit collections, metadata filters, hybrid search, persistence or multiple applications. A small single-user prototype may not need another service. Self-hosted Qdrant starts without authentication or encryption by default, so never expose its ports unchanged. Apply the retrieval security checklist.
Vector searchLocal or cloudSecure explicitly
Chroma
Chroma offers an approachable local or self-hosted retrieval store, with a managed route available separately. Define metadata filters, tenancy, authentication, backups and an export path before relying on it—the vector index is rebuildable infrastructure, not the source of truth. See the reference architecture.
Local or self-hostedVector searchExport plan
Chroma documentation →Weaviate + Milvus
Both support vector search, metadata filtering and production deployments across self-managed and hosted routes. They introduce a real database service: size indexes, define tenancy, secure every endpoint and test restore and re-embedding before relying on either. Follow the production build gates.
Vector databaseSelf-hosted optionsOperate deliberately
Neo4j + FalkorDB
Add a graph only when relationships and multi-hop traversals are part of the acceptance test; PostgreSQL is enough for an appointment queue. Neo4j Community is GPL-3.0. FalkorDB targets GraphRAG but uses SSPL, so it is source-available rather than OSI open source. GraphAware is a commercial graph-services layer, not a required database component. Use the GraphRAG decision gate.
Knowledge graphsGraphRAGLicense check
Mem0 + Graphiti
Mem0 provides an open-source memory layer; Graphiti is the self-hosted temporal context-graph engine behind hosted Zep. Neither makes extracted memories authoritative: preserve raw records and consent rules, namespace every user and agent, and test correction, deletion, provenance and stale-memory behavior before using recalled facts in decisions.
Long-term memorySelf-hostedRetraction test
Cognee
Cognee is a self-hosted memory framework that ingests supplied data into graph and vector structures and exposes add, cognify, search and memory workflows. It can use local databases and OpenAI-compatible local model endpoints; configure both generation and embeddings explicitly, keep source records outside the derived graph, and test deletion and re-indexing before production use.
Self-hostedGraph + vectorSource records required
Cognee repository →Letta + Redis
Letta builds stateful agents around persistent memory and tools. Redis can hold fast operational state and provide search primitives, but it does not define memory policy for you. Separate conversation state, durable facts and source records; give each a retention and deletion path.
Agent memoryStructured stateRetention required
Node-RED + n8n
Both provide visual triggers, API calls and approvals. Node-RED is Apache-2.0 and fits a strict OSS stack. n8n has the larger packaged integration surface but uses a Sustainable Use fair-code license; keep that distinction visible in procurement.
VisualConnectorsDifferent licenses
Zapier + Make + Pipedream
Zapier and Make emphasize visual SaaS automation; Pipedream combines hosted connectors with code steps. They are useful for bounded integrations and approvals, but credentials and payloads cross a managed control plane—review data handling, retries, run history and exportability.
ManagedConnectorsVisual + code
Temporal
Use Temporal when timers, callbacks, retries and state must survive process restarts for hours or days. Its MIT-licensed server and .NET SDK fit the appointment workflow: deterministic workflow code owns state while activities perform calendar, PBX and database I/O.
MITDurable state.NET SDK
Prefect + Apache Airflow + Kestra
Use these for scheduled data flows, backfills and observable batch dependencies. Prefect and Airflow are Python-centered; Kestra defines workflows declaratively and spans more runtimes. They can launch preparation or reporting, but long-lived business conversations may be clearer in Temporal or application code.
SchedulesBackfillsObservable runs
Dify + Flowise
Visual options for retrieval, model routing and tool-using flows. Useful for prototypes when evaluation and export remain part of the design.
VisualSelf-hostablePrototype
Move, transform and serve data
Meltano + dlt
Code-first, open-source routes for moving API and database data into an analytical store without adopting a managed connector platform. Use them when the connector set is modest and versioned pipelines matter more than a large visual catalog.
OSSELTCode-first
Airbyte + Fivetran
Airbyte offers a broad self-hosted connector catalog but its main platform uses Elastic License 2.0, not an OSI license. Fivetran is a managed proprietary service. Compare them for connector coverage and operations—not as equivalents to a true OSS stack.
Many connectorsELTLicense boundary
dbt Core
Version SQL transformations, tests and documentation after data lands in an analytical database. dbt is not an ingestion tool, operational scheduler or appointment transaction engine; use it for trusted reporting models.
SQLTestsLineage
dbt Core →Apache Iceberg + Delta Lake
Transactional table formats for large analytical data lakes, schema evolution and multiple compute engines. They become relevant when call/event history reaches lake scale; they do not replace PostgreSQL for live slot locks.
Apache-2.0Data lakeLarge scale
Apache Superset + Dashjoin
Superset is an Apache-licensed BI and data-exploration layer for outcome dashboards. Dashjoin is an AGPL low-code platform that can put forms and actions over existing sources. Neither should become the authoritative booking state.
BILow-code UIRead-mostly
Metabase
An AGPL-licensed, self-hostable service for queries, dashboards and self-service exploration when a simpler BI interface fits better than Superset. Run the Open Source Edition as a Java JAR or Docker container with a production application database. Self-hosting keeps Metabase the company out of your data unless you opt in to anonymized usage statistics, but it does not make connected data sources local or safe by default. Use a dedicated least-privileged, read-only database user, keep results and dashboard access behind authenticated groups, and keep transactional state in its application service. Fine-grained row and column controls require a self-hosted Pro or Enterprise plan.
AGPL-3.0JAR or DockerLeast-privilege data access
Build operational intelligence systems
JanusGraph + TinkerPop + Cassandra
JanusGraph supplies a distributed property-graph layer, TinkerPop defines the Gremlin traversal API and Cassandra can provide the storage backend. Use this combination only when relationship-heavy queries and scale justify three moving parts; keep raw and curated records outside the graph so it remains a rebuildable serving model.
Open sourceDistributed graphOperational burden
Apache Kafka + Apache NiFi
Kafka fits durable event streams and replay; NiFi fits visual routing, provenance and source-to-destination controls. A nightly file does not need either. Name which service owns retry and deduplication, retain raw inputs and test what happens when the source sends late, duplicated or malformed records.
Apache-2.0StreamingProvenance
Apache Iceberg + S3-compatible storage
Iceberg adds schema evolution, partition management and snapshots over object storage. Default new self-hosted builds to SeaweedFS; consider Garage for lightweight multi-site replication or Ceph RGW when Ceph already has an operations team. MinIO Community is no longer a default: its upstream repository is archived and its distribution is source-only. Pin the catalog, engine and store versions, then prove compatibility and a consistent restore.
SeaweedFS defaultGarage + Ceph alternativesMinIO migration only
Apache Spark + Trino + DuckDB
Spark handles distributed transforms, Trino provides interactive SQL across services and DuckDB is often enough for the first single-machine proof. Start with DuckDB, preserve standard table formats and add distributed engines only when measured data size, concurrency or latency demands them.
SQL + computeScale laterOpen formats
Apache Solr
Solr provides text search, faceting and geospatial queries and can act as a JanusGraph mixed index. Treat the index as derived and rebuildable, apply document-level authorization before results leave the service and test stale-index behavior rather than letting search become an accidental system of record.
Apache-2.0SearchDerived index
Solr reference guide →GeoTrellis + CesiumJS + MapLibre
GeoTrellis handles tiled raster and terrain work on Scala/Spark; CesiumJS renders a 3D globe; MapLibre renders conventional 2D and 2.5D maps. These are processing and presentation components, not authoritative data. Keep source time, accuracy and attribution visible in the map and request only the viewport the operator may see.
Open source2D + 3DGeospatial
LocStat + Apache Superset
LocStat is useful as a Palantir-style interface reference; verify code availability, support and licence before making it a dependency. Superset is an Apache-licensed analytical dashboard, not a transactional workflow engine. Borrow interaction patterns, but keep approvals and case state in a maintained application service.
Interface referenceBIVerify licence
Apache Airflow + Prefect
Either can schedule, retry and observe batch ingestion and ontology projection. Choose one owner for each retry, make tasks idempotent and keep long-lived human cases in application state or a durable workflow engine. A scheduler succeeding does not prove that the resulting data is correct.
PythonSchedulesOne retry owner
Apache Ranger + Open Policy Agent
Ranger centralizes policies and audit for supported data services; OPA evaluates policy as code in applications and infrastructure. Neither creates identities, secrets or correct object-level permissions automatically. Test denied rows, fields, actions and exports with the exact service versions in production.
AuthorizationAuditTest denies
vSphere Supervisor + Cloud Director CSE + vSphere CSI
Use ordinary vSphere VMs for a small stateful deployment unless the datacentre already operates Kubernetes. Supervisor and Cloud Director CSE add cluster or tenant control planes; the CSI driver maps persistent volumes to vSphere storage policies. Tenant boundaries do not replace application authorization, and VM snapshots do not replace application-consistent backup.
VMwarePrivate deploymentRecovery test
OpenFoundry + Osiris interface references
These repositories can make ontology, lineage, map, layer and event-card choices tangible. Treat them as workshop material until their licences, maintenance, identity model and data boundaries pass review. Reimplement only the operator journey your own services and acceptance tests require.
PrototypeInterface studyNot a platform proof
Google OR-Tools
Use established solvers for vehicle routing, assignment, scheduling and other constrained decisions before asking a language model to improvise. Define hard constraints separately from preferences, preserve the solver inputs and compare feasible solutions against the current operational baseline.
Apache-2.0OptimisationDeterministic baseline
OR-Tools documentation →OpenSky Network
A research-oriented REST API for live aircraft state vectors, tracks and limited flight history. Coverage and identity are receiver-dependent, quotas apply and OpenSky now requires a written agreement for operational or commercial REST API use. Use licensed snapshots as realistic fixtures; use an approved aviation feed or your own authorized receivers when decisions depend on completeness.
Live + historicalWritten licence for operationsNot authoritative
ADS-B Exchange
The commercial API exposes changing aircraft positions and a military-filtered endpoint. It is useful as a supplementary or development feed, not proof of identity, intent or complete coverage. Budget for subscription and rate limits, preserve source timestamps and reconcile against the authority responsible for the air picture.
Commercial APILive aircraftSupplement only
CelesTrak GP data
CelesTrak publishes current general-perturbations element sets in TLE, OMM JSON, XML, KVN and CSV formats. Rendered paths are propagated estimates from an element epoch—not continuous live sensor positions. Pin dated fixtures for development and surface element age, propagation time and uncertainty in operational views.
Public dataOrbit propagationShow data age
OpenStreetMap data
OpenStreetMap can supply roads and mapped features for routing, simulation and basemaps; it does not supply live vehicle flow. Attribute contributors and respect the ODbL. The community tile service is best-effort and forbids bulk or offline use, so production systems should self-host approved extracts and tiles or use a provider with a contract.
Open dataAttribution requiredSelf-host production tiles
City of Austin Traffic Cameras
A public-domain dataset with camera identifiers, approximate locations, status and—where publication is allowed—a link to the latest screenshot on a five-minute cadence. It is useful for a public-feed prototype and outage fixtures. The city says footage is not retained and the locations are not suitable for legal, engineering or surveying use.
Public domainFive-minute screenshotsApproximate locations
Official dataset →Frigate
An MIT-licensed, self-hosted NVR that performs real-time object detection locally for IP cameras, retains recordings according to detected objects and integrates with Home Assistant and MQTT. Use a supported GPU or AI accelerator as the project recommends, and keep its unauthenticated internal API off exposed networks; use the authenticated endpoint or a carefully configured reverse proxy. “Local” still leaves sensitive footage, camera credentials and retention policy for you to secure.
MITLocal object detectionAccelerator recommended
Google Photorealistic 3D Tiles
A high-resolution textured 3D basemap that can be rendered with CesiumJS. It requires billing, an API key and visible attribution. Google's policies restrict prefetching, storage, offline use, extraction and machine analysis, so keep every operational layer independent and retain a fallback to government terrain, imagery or 3D city models.
Hosted APICommercial termsBasemap only
PostGIS + GeoServer + QGIS
PostGIS stores governed spatial records in PostgreSQL, GeoServer publishes standards-based services and QGIS supports desktop editing, analysis and review. This is the airtight fallback for an organisation that already owns authoritative GIS data: keep identity and workflow state in maintained services, issue read-only layers to the browser and test disconnected operation and restore.
Self-hostedOGC servicesAuthoritative data
Build software with agents
RTK
An Apache-2.0 Rust proxy that rewrites supported shell commands and filters, groups, deduplicates or truncates their output before it reaches a coding agent. The project estimates 60–90% token savings on common commands; measure that on your workload. Retain a route to the raw command output because compression can hide the line that matters during diagnosis.
Rust binaryAgent hooksKeep raw output
RTK repository →Ralph
An MIT-licensed shell loop that turns a PRD into prioritized stories, launches a fresh supported coding-agent instance per story, runs checks and commits passing work until a limit is reached. It is a useful scaffold, not proof that unattended agents finish correctly, and the selected agent keeps its own model and data boundary. Use a disposable branch, cap iterations and review every commit.
External coding agentFresh contextsBound the loop
Ralph repository →Caveman
An MIT-licensed skill that asks coding agents to answer in terse “caveman” language while preserving code and commands. Its 65% output-token claim comes from the project and does not reduce input context. Test whether terseness drops rationale, warnings or accessibility for your team; avoid it where a complete handoff matters more than output cost.
Many agentsOutput onlyVendor benchmark
Caveman repository →Codex CLI
OpenAI's open-source terminal coding agent can run against Ollama or LM Studio with --oss; --local-provider selects the provider and oss_provider sets a default. Local inference is not a containment boundary: the agent can still read, edit and run commands. Keep sandboxing and approvals enabled, scope files and credentials, choose a model that handles tools reliably and verify the result with tests and review.
Open sourceOllama + LM StudioSandbox + approvals
Codex Router
An independent MIT-licensed beta that adapts selected external providers to the Responses API and merges their models into the Codex App and CLI picker through a loopback service. It writes marked Codex configuration blocks, installs a per-user background service and stores provider credentials in protected local files. Review a tagged installer instead of piping the moving default branch, keep its config backup and rollback path, run the included doctor and re-test after Codex updates. Provider billing, data handling and terms still apply, and routed models do not become OpenAI-supported Codex models.
MITBetaConfig + credential boundary
Code Buddy
An MIT-licensed coding agent with a terminal UI, Electron-based Cowork desktop app, HTTP/WebSocket server, local Ollama support, hosted model providers, MCP and a broad tool surface. Local inference does not contain file, shell, browser or connector access: begin with narrow permissions, isolate credentials and repositories, keep optional background and self-improvement features off until tested, and review generated changes and external actions.
MITCLI + desktopLocal + hosted models
Code Buddy repository →JCode
An MIT-licensed terminal coding agent with hosted and local model routes, MCP, semantic memory, persistent background work and coordinated swarms. The project publishes its own optimization benchmark plus Linux PSS and startup comparisons; treat those as project-run evidence and reproduce them with pinned versions. Scope files, shell, network and credentials, and set retention rules so recalled memory does not preserve stale decisions or sensitive material.
RustLocal + hosted modelsMemory + swarms
OpenCommit
An MIT-licensed CLI that generates commit messages from staged changes through a hosted model or local Ollama. A remote provider may receive the diff, and the default flow can stage files for you. Stage deliberately, exclude secrets, inspect the final diff and edit the message so it describes the actual change rather than the model's guess.
Git CLIHosted or local modelReview before commit
OpenCommit repository →GitButler
A Git-based desktop and CLI workflow for organizing coding-agent changes into parallel or stacked branches, selected hunks and reviewable commits in one working directory. Its agent setup can install version-control instructions, but those instructions are not access controls; keep repository permissions and release approval separate. Workspace mode also changes branch semantics, so follow GitButler's commands instead of mixing in Git index, checkout or branch operations. The source uses a Fair Source licence that converts to MIT after two years and restricts competing products during that period.
Desktop + CLIParallel branchesFair Source
Buzz
Block's Apache-2.0 workspace puts people and model-agnostic agents—including Codex, Claude Code and Goose harnesses—into shared channels around conversation, workflows and code. Events live on a Nostr relay that can be hosted by Block or self-hosted. Treat it as a collaboration and audit layer, not a sandbox: each underlying agent still needs narrowly scoped repositories, tools, network access and credentials. Buzz is pre-1.0, so check the current release and the project's works-now / in-progress table before standardizing a team workflow.
Apache-2.0Hosted + self-hostedHuman + multi-agent
Claude Squad
An AGPL-3.0 terminal interface for running Claude Code, Codex, Gemini, Aider and other agents in separate Git worktrees. Worktrees reduce merge collisions but are not process, network or credential sandboxes; its auto-accept mode further expands authority. Review changes before applying them and isolate secrets and external accounts separately.
TUIGit worktreesAGPL-3.0
Claude Squad repository →ParallelCode
A free desktop Git worktree manager for opening Cursor, Claude Code, Copilot and other local-repository agents on separate branches. It is the focused graphical choice when parallel worktrees—not a full task platform—are the requirement. Worktrees separate files and branches, not processes, ports, networks or credentials, so coordinate shared services and review every merge.
Desktop GUIGit worktreesFree
Windows Terminal
Microsoft's MIT-licensed terminal host puts PowerShell, Command Prompt, WSL and SSH profiles into configurable tabs and split panes. It is a clean baseline for watching several Windows-side agent CLIs, but it does not add agent status, task routing, worktrees or review workflow. Use scripted wt layouts for repeatability and pair it with Git or a session manager when branches and resumable state matter.
WindowsTabs + panesMIT
Nimbalyst
A free, MIT-licensed desktop workspace for parallel Codex and Claude Code sessions, Kanban task management, worktrees, terminals and visual editing of documents, diagrams, mockups and code. It is the strongest fit here when non-terminal review and a mobile companion matter. Core content stays in local files, but anonymous analytics are enabled unless you opt out and mobile collaboration uses a separately operated sync service; verify both boundaries before using sensitive repositories.
Visual workspaceCodex + Claude CodeMIT
Orca
An MIT-licensed desktop workspace for running Codex, Claude Code, OpenCode and other CLI agents side by side in Git worktrees, with terminals, editing, per-worktree browsers, diff review and SSH. Mobile role: its beta iPhone and Android companion is a phone remote control for desktop-owned sessions: watch status and terminal output, browse files, reply or dictate, review changes and reconnect over LAN or Tailscale. Orca's current docs say supported agents launch with their full-autonomy flags pre-applied by default. Worktrees separate branches, not process, network or credential authority, so add real containment or change that launch policy before treating a pocket approval as a safety boundary. Packaged builds also send opt-out anonymous usage telemetry to PostHog's US region.
Phone companion betaDesktop source of truthMIT
Kandev
An AGPL-3.0, self-hostable control plane with Kanban and pipeline views, worktrees, an integrated editor, terminal and review flow. Its broad agent registry includes Claude Code, Codex, Copilot, Gemini, OpenCode, Kimi and Grok, while executors can run locally, in Docker, over SSH or in a cloud environment. That breadth brings more operational surface: pin agent adapters, scope executor secrets and treat container or remote-host policy as a separate security boundary.
Many agent vendorsLocal + remote executionAGPL-3.0
Agent Deck
An MIT-licensed TUI that adds agent-aware running, waiting and completed states to tmux, plus groups, search, session forking, Git worktrees, cost tracking and local or remote sessions. It is the lightweight terminal-native choice and installs through Homebrew, Go or a release script; Windows use runs through WSL. Session organization is not command isolation, so keep each underlying agent's approvals, sandbox and credential scope intact.
TUI + tmuxHomebrewMIT
CCManager
An MIT-licensed terminal manager for parallel Claude Code, Codex, Gemini, Cursor Agent, Copilot, Cline, OpenCode and Kimi CLI sessions across projects and Git worktrees. Busy, waiting and idle indicators make it a straightforward choice when status visibility and broad CLI support matter more than a graphical workspace. Its worktree merge and delete actions change Git state, and experimental auto-approval expands agent authority; keep both deliberate and recoverable.
TUIKimi + major CLIsMIT
herdr
An Apache-2.0 Rust terminal runtime that keeps coding-agent panes alive in a background server, marks them working, blocked or idle and supports detach, reattach, SSH access and an agent-facing socket API. It runs existing tools such as Codex, Claude Code, Cursor and OpenCode rather than replacing them. Persistence and status tracking are not containment: each pane still inherits the filesystem, network and credentials available to its process, so keep agent approvals and sandboxing in place.
Rust binaryPersistent sessionsApache-2.0
herdr repository →Headroom
An Apache-2.0 local-first library, proxy and MCP server that compresses tool output, files, logs and retrieval chunks before model use. The project reports 15–20% savings for coding agents and 60–95% for JSON, with reversible paths. Benchmark answer quality as well as token count, keep raw artifacts and bypass compression for forensic or exact-text work.
Library + proxy + MCPLocal-firstKeep originals
Headroom repository →OpenHands Agent Canvas
A beta, MIT-licensed interface for running OpenHands, Claude Code, Codex, Gemini and ACP-compatible agents across local, remote or cloud backends. Local UI does not mean local inference or isolated execution: inspect the selected backend, mounts, network, provider credentials and automation triggers before leaving it always on.
BetaMany agent backendsLocal + remote
Agent Canvas repository →Visual Studio Code
VS Code's built-in language-model picker can connect chat and local-agent workflows to Ollama or a custom OpenAI-compatible endpoint without a Copilot plan or GitHub sign-in. Native BYOK does not provide standard inline completions, semantic search or embedding-backed features. On DGX Spark, use NVIDIA's guide for a direct ARM64 installation or remote access through NVIDIA Sync.
Local or remoteBYOKChat + agents
Continue
An open-source assistant for chat, edits, agents and role-specific autocomplete. It can discover Ollama models locally or use a remote Ollama address, which fits a Spark reached through a private network or SSH tunnel. Pick models by role: a good chat model is not automatically a good fill-in-the-middle completion model.
Local-capableIDE extensionRole-specific models
Roo Code
A model-flexible coding agent that can use Ollama, LM Studio or another OpenAI-compatible endpoint. Local operation still depends on a model with reliable tool calling and enough context; start with approvals enabled and test file, terminal and browser boundaries before increasing autonomy.
Local-capableAgenticApproval controls
JetBrains AI Assistant
IntelliJ IDEA, Rider and the other supported JetBrains IDEs can connect AI Assistant to Ollama, LM Studio or a private OpenAI-compatible endpoint. Local models can cover chat and assigned editor tasks, but local-model MCP tools, next-edit suggestions and some proprietary completion features remain unavailable.
Local or remoteIDE-nativeFeature caveats
Custom and local models →Zed
Zed can connect its Agent and Inline Assistant directly to Ollama, LM Studio, llama.cpp or a local OpenAI-compatible server. Remote endpoint URLs are supported, while local or self-hosted edit prediction has a separate configuration path.
Native editorLocal endpointsAgent + inline
Local-model guide →OpenCode + OpenSpec
A provider-flexible coding agent paired with a durable, reviewable agreement about what to build and what stays out of scope.
TerminalProvider-flexibleSpec-driven
Aider
A terminal coding assistant built around repository edits and version control. Use a clean branch and keep tests as the acceptance boundary.
TerminalGit-awareMultiple providers
Official documentation →Cline
An open-source coding agent for IDEs and the terminal with reviewable diffs and human-in-the-loop approval. Keep its repository scope narrow and review every command or enablement choice.
IDE + CLIModel-flexibleReviewable diffs
Project repository →Pi Agent Harness
The MIT-licensed earendil-works/pi project packages a multi-provider LLM API, tool-calling agent core, terminal UI and extensible coding-agent CLI. Pi explicitly does not provide a built-in permission system, so run it with only the filesystem, process, network and credentials it should have—prefer a tested container, micro-VM or policy sandbox for consequential work.
TypeScriptMulti-providerExternal sandbox required
Prime Agent
An MIT-licensed coding and research agent built around a persistent Python environment, programmatic subagents and durable harness state that can retain reviewed memories, prompts, skills and subagent specifications. Background sessions, goals, schedules and bounded autonomous runs suit long work. It executes model-generated Python and project commands with the user’s permissions—not inside a security sandbox—so use a disposable worktree or external containment and inspect every retained refinement.
MITTerminal + background daemonExternal sandbox required
Prime Agent repository →Nanocoder
An MIT-licensed TypeScript terminal agent with Ollama and other local-server routes, cloud providers, MCP, skills, subagents, checkpoints and scheduled or event-driven runs. “Local-first” describes where the selected model can run, not a permission boundary: file and shell tools still act with the process’s authority. Begin in normal or plan mode, isolate the repository and credentials, and measure its exploration cost on your own tasks.
Terminal + VS CodeLocal + cloud modelsScope tool authority
LibreCode
A .NET 10 and Avalonia code editor with Ollama integration, terminal and model browsing for Windows, Linux and macOS. Its repository is public and local-first, but the custom license restricts redistribution, forks and SaaS use—treat it as source-available, not permissively open source.
.NET + AvaloniaOllamaCustom license
Context7
Feeds current, version-specific library documentation into coding agents through a CLI skill or MCP server. Pin the library ID and version, and still open the primary documentation for consequential changes; a retrieved snippet is context, not a compatibility test. Hosted use sends the documentation query to Context7.
CLI or MCPCurrent docsCoding agents
opensrc
An Apache-2.0 Rust CLI that gives coding agents searchable source for npm, PyPI and crates.io packages plus GitHub, GitLab and Bitbucket repositories. It resolves an installed or requested version, shallow-clones the matching source and caches it locally so ordinary tools such as rg can inspect implementation, tests and examples. Pin the dependency version and verify the resolved tag: code is the strongest evidence of implementation behavior, while official release, support and security guidance still define the supported contract.
Apache-2.0Rust CLIPackage + repository source
Fine-tune, evaluate and reproduce
Netron
An MIT-licensed desktop, browser and Python viewer for ONNX, TensorFlow, PyTorch, Core ML, OpenVINO, Safetensors and many other model formats. Use it to inspect graph structure, operators, tensor shapes and metadata before conversion or deployment. A readable graph is not proof that a checkpoint is safe, correctly licensed or numerically equivalent after export; keep provenance, hash files and run format-specific validation.
MITDesktop + browser + PythonInspect, then validate
iFixAI
An Apache-2.0 CLI and agent skill that runs 45 inspections, with 32 core checks contributing to an A–F score and optional independent model judges. The project explicitly says the result is not certification or a safety guarantee. Version fixtures and judges, inspect false positives and unsupported capabilities, budget judge calls and disable disclosed telemetry if policy requires it.
Agent evalsCI-friendlyNot certification
iFixAI repository →Hugging Face PEFT + TRL
PEFT provides adapters such as LoRA; TRL provides supervised and preference-training loops. This is the flexible foundation when your team wants code-level control over datasets, trainers and evaluation.
LoRASFT + DPOCode-first
Axolotl
Packages full, LoRA, QLoRA, preference and reinforcement-learning runs into reusable YAML configuration, with distributed and multi-GPU routes. Pin the repository revision, base model, dataset, chat template and dependency environment; a portable recipe is not a guarantee that another GPU topology will reproduce the same speed or result.
Config-drivenDistributed trainingAdapters + full tuning
Axolotl documentation →Unsloth
Unsloth Desktop is a beta, cross-platform app for running and training local text, audio, image and video models, with RAG, MCP, code execution, export and OpenAI- and Anthropic-compatible endpoints that can connect Codex or Claude Code. Its Core package is Apache-2.0; the Studio UI is AGPL-3.0. Unsloth reports up to 2× faster supported LLM training with up to 70% less VRAM, but treat that as vendor evidence and reproduce it on the exact model and hardware. Support differs across CPU, NVIDIA, AMD, Intel and Apple backends. Keep it on loopback first: web search, cloud providers, MCP servers and Cloudflare remote access each change the data or authority boundary, and anyone with the tunnel link and API key can run code.
Desktop + CoreLocal multimodal + trainingBeta · dual license
Microsoft Agent Lightning
Agent Lightning is an MIT-licensed training layer that turns agent trajectories into inputs for reinforcement learning, supervised fine-tuning or prompt optimization without binding the agent to one framework. The trainer, inference engine and compute remain your responsibility: version rewards and learned artifacts, retain untouched holdouts and reject gains that disappear under repeat runs.
Self-deployedFramework-agnosticRL + SFT
Agent Lightning repository →MLflow
Track run configuration, artifacts and metrics, and compare an adapter with its baseline and holdout. Keep human task judgments alongside automated scores; a model judging another model is not independent proof.
Run trackingArtifactsEvaluation
GenAI evaluation docs →Arize Phoenix + W&B Weave
Phoenix and Weave trace model and agent calls, organize datasets and run evaluations. Instrumentation can capture prompts, retrieved documents and tool arguments, so define redaction, sampling, access and retention before sending production traces anywhere.
TracingDatasetsEvals
Langfuse + DeepEval
Langfuse self-hosts tracing, prompt management, datasets and evaluations, but production deployment brings Postgres, ClickHouse, Redis or Valkey and object storage. DeepEval is an Apache-2.0, pytest-style evaluation framework that can run in local development and CI. Local tooling does not guarantee local judging: configure the evaluator model deliberately and redact sensitive trace fields before capture.
Self-hostedTracing + evalsCI tests
TruLens + Ragas
TruLens combines tracing with feedback functions for agent applications; Ragas provides evaluation workflows and metrics for RAG and other LLM systems. Calibrate automated scores against human judgments and keep a stable, versioned test set—metric names are not evidence by themselves. Use the RAG evaluation plan.
RAG evalsAgent evalsHuman calibration
Promptfoo + Helicone
Promptfoo runs repeatable model comparisons, assertions and adversarial tests; Helicone captures and analyzes LLM requests through an observability layer. Keep golden cases in version control, block releases on meaningful thresholds and avoid treating traffic dashboards as correctness tests.
Regression testsRed teamingRequest traces
garak + PyRIT
Apache-2.0 garak probes models for prompt injection, jailbreaks, data leakage, misinformation, toxicity and other failure modes; MIT-licensed PyRIT orchestrates multi-turn attacks, converters, scorers and result memory for structured security exercises. Run only against systems you own or are authorized to test, cap cost and concurrency, isolate credentials, and handle prompts, responses and JSONL logs as potentially harmful security fixtures. Community attack corpora such as L1B3RT4S can broaden coverage, but they are uncurated inputs—not a safety standard or a pass/fail oracle.
Attack automationMany model routesAuthorized targets only
Heretic
An AGPL-3.0 command-line tool that automatically finds and suppresses refusal directions in supported open-weight transformer models, then offers comparison and evaluation routes. That makes it useful for interpretability and controlled robustness research—and capable of deliberately removing safety behavior. Use only in an isolated, authorized lab; preserve the original checkpoint, block deployment and sharing by default, scan generated artifacts, and compare helpfulness, over-refusal and harmful compliance rather than celebrating a lower refusal count.
Directional ablationDual-useLab isolation
Tensor Trust + Gandalf
Tensor Trust is an open research game where participants write defenses and attack other prompt-protected accounts; Gandalf is a guided prompt-injection challenge that becomes harder across levels. They make failure modes tangible for awareness training, but neither simulates tool permissions, retrieval poisoning or a production incident. Tensor Trust says submitted text will be released for research, so never enter secrets, personal data or proprietary prompts.
Hands-on learningAttack + defensePublic submissions
Run durable assistants and custom systems
Cloudflare OS
An Apache-2.0 early-access workspace for agent chat and AI-built, shareable “Gadgets” on Cloudflare Workers. Capability-scoped Gatekeepers mediate external resources, log actions and queue side-effect approvals. The quick local route is explicitly non-production, while documented deployment on your own workerd server is still forthcoming; audit each OAuth connector, sandbox boundary, simulated approval result and data-residency path before organizational use.
Apache-2.0Workers + workerdEarly access
Rivet agentOS
An Apache-2.0 runtime that executes supported coding agents inside lightweight isolated Linux environments with deny-by-default filesystem, network and process permissions. Its cold-start and cost comparisons are project benchmarks. Test required syscalls, process behavior and deny paths on your workload; a lightweight VM boundary still needs patched hosts, secrets brokering and resource limits.
Linux environmentDeny by defaultVerify compatibility
agentOS repository →MEMANTO
An MIT-licensed local memory service with remember, recall and answer operations for several coding-agent clients, without a hosted backend or vector database. Persistent memory can preserve mistakes, secrets and stale decisions as easily as useful context. Scope what is ingested, test retrieval on a known set and keep export, deletion, backup and retention controls visible.
LocalPersistent memoryNo vector database
MEMANTO repository →Microsoft UFO
UFO is an MIT-licensed Microsoft research project for Windows desktop automation and cross-device orchestration. The code runs on infrastructure you control, while model endpoints are configured separately. GUI control carries broad authority: begin with non-critical accounts, isolate credentials and files, require approval for consequential actions and measure recovery from focus changes and partial execution.
Self-deployedWindows + devicesResearch project
Microsoft UFO repository →Goose
An Apache-2.0 desktop app, CLI and API for coding, research and automation across many hosted and local model providers, with MCP extensions, reusable recipes and ACP support. It includes tool permissions, prompt-injection checks and optional adversary review, but those controls do not make arbitrary extensions or model output trustworthy. Begin with a narrow directory, an extension allowlist and approval-required tools; inspect provider, subscription and telemetry paths separately.
Desktop + CLI + APIMCP + ACPPermission controls
Hermes Agent
A persistent assistant that can learn reusable skills and run on Linux, macOS, Windows or a remote server. Mobile role: talk to the same home- or cloud-hosted agent through Telegram, Discord, Slack, WhatsApp or another configured channel. The messaging provider remains part of the communications path.
Phone messagingRemote-firstSkills
OpenClaw + companion apps
OpenClaw keeps one gateway and state store on the always-on host, while Windows, macOS, Linux, mobile and messaging clients act as operators or narrowly scoped nodes. Mobile role: its iPhone and Android companions connect to that gateway; the phone does not host the gateway or silently reduce its tool authority.
iPhone + AndroidAlways-on hostRemote gateway
TripleBits Apprentice
A desktop route for creating scheduled agents with chosen models, memory, budgets, channels and tools. TripleBits says Apprentice runs agents in isolated local Docker containers, offers command and website controls and keeps run logs locally; treat those as vendor claims and test filesystem mounts, network denial, credential brokering and budget stops before sensitive work.
Local-firstSchedules + channelsVerify controls
TripleBits Apprentice →NVIDIA OpenShell
A policy and sandbox runtime for autonomous agents on Linux, macOS and Windows through WSL 2. Use it to bound files, networking and execution around an agent or tool runner; still verify the selected backend and test the deny path instead of treating “sandboxed” as a blanket guarantee.
PolicySandboxCross-platform
Official repository →LangChain + LangGraph
LangChain supplies model and tool integrations; LangGraph adds explicit stateful agent graphs, checkpointing and human-in-the-loop control. They orchestrate agent reasoning—not telecom media, distributed transactions or durable business state by themselves.
Code-firstAgent statePython + TypeScript
Microsoft Agent Framework + Governance Toolkit
Use Microsoft Agent Framework for new .NET or Python agents and graph-based workflows; Microsoft describes it as the successor to AutoGen and Semantic Kernel, and its current orchestration patterns cover the Magentic route without a separate legacy Magentic-One deployment. Pair it with the framework-agnostic Agent Governance Toolkit for policy enforcement, agent identity, sandboxing and tamper-evident audit records. Pin versions and test deny paths before production.
.NET + PythonWorkflowsGovernance controls
Operate the stack
OpenTelemetry + Prometheus + Grafana + Loki
Trace one workflow across service boundaries, record bounded metrics, build dashboards and aggregate logs. Avoid putting transcripts, phone numbers, prompts or appointment details into metric labels and unbounded log streams.
TracingMetricsLogs + alerting
Keycloak + OpenBao
Use Keycloak for operator identity and roles; use OpenBao for SIP, calendar and database credentials with short leases or rotation where supported. Neither replaces host hardening or application-level authorization.
OIDCSecretsRotation
Microsoft Purview
Microsoft's portfolio spans data cataloging and governance, information protection, data loss prevention, audit, lifecycle, eDiscovery and related compliance controls. In an AI teammate, use Purview to classify and govern data around the workflow; it does not replace per-tool authorization, short-lived credentials or runtime containment.
Microsoft ecosystemData controlsAudit + compliance
NVIDIA NeMo Guardrails + Guardrails AI
NeMo Guardrails adds configurable input, dialog, retrieval, execution and output rails around LLM applications. Guardrails AI focuses on validators and structured, checked outputs. Use either as explicit policy code with tested allow and deny cases—the framework name is not a safety guarantee.
Policy checksValidationTest deny paths
Microsoft Presidio
Presidio detects and anonymizes personally identifiable information across text, images and structured data. Detection is probabilistic and domain-dependent: add organization-specific recognizers, measure false negatives and keep access control and minimization in place even after redaction.
PIISelf-hostableCustom recognizers
Microsoft Presidio →Prompt Security + Protect AI
Prompt Security focuses on discovering and protecting generative-AI use and applications. Protect AI covers AI security across models, artifacts and the software supply chain. Evaluate the exact product or open-source scanner you intend to deploy; neither replaces least privilege, sandboxing or incident response.
Application securityModel supply chainCommercial platforms
SafePrompt + Lakera Guard + LLM Guard
SafePrompt is the low-friction hosted choice for most developers: one provider-neutral HTTP call and a free starting tier. Its own April 2026 comparison reports most checks under 100 ms and accuracy above 95%; treat those as vendor claims and validate them against your traffic and attack set. Choose Lakera Guard when current SOC 2 evidence and enterprise support are procurement gates, or MIT-licensed LLM Guard when prompts must stay inside your infrastructure and you can operate and tune Python scanners. Detection is one layer—not a replacement for least-privilege tools, policy enforcement, output validation and incident logs.
Hosted defaultEnterprise optionSelf-hosted option
NATS + RabbitMQ
Add a broker when independently deployed services need backpressure, fan-out or replayable delivery. A single Temporal deployment plus PostgreSQL may not need one; do not introduce a second retry system until ownership and deduplication are explicit.
MessagingBackpressureOptional
No tool matches this topic and search. Try a broader term or choose all.
Selection test
Trial the boundary, not only the happy path.
A useful tool survives your real files, denial cases, export needs and maintenance budget.
Test ten representative cases against a manual baseline and a named acceptance threshold.
Write down what leaves the device, which identities and tools it can use, and where traces remain.
Exercise denied files, missing credentials, hostile content, bad output and recovery after interruption.
Export data and configuration, reproduce one result and estimate the work to replace the tool.
Coding-agent overhead
Efficiency: wall-clock and memory
Two narrow studies expose costs hidden by capability scores. They use different methods and workloads, so read each on its own terms—not as one combined ranking.
Hold DeepSeek V4 Flash fixed; measure the wrapper.
The study ran the same eight focused repository bug fixes with the same local model and grading method. Pi, OpenCode and Claude Code have 24 timed runs each; Nanocoder has 54 timed runs in the published wall-clock series. Dots are individual runs and diamonds are task-weighted averages.
Axis range
Pi2.1m average24 runs · 4 tools
OpenCode3.1m average24 runs · 10 tools
Claude Code8.0m average24 runs · 27 tools
Nanocoder5.2m average54 timed runs · 15 tools
Add another session; measure proportional RAM.
JCode's corrected Linux rerun measures the slope from one to ten active clients. Bars show approximate extra proportional set size (PSS) per added session, which accounts for shared memory proportionally. Lower is better.
JCode · embeddings off~9.9 MB · baseline
JCode~10.4 MB · 1.1×
Codex CLI~21.6 MB · 2.2×
Pi~76.5 MB · 7.7×
Antigravity CLI~86.4 MB · 8.7×
Cursor Agent~157.5 MB · 15.9×
GitHub Copilot CLI~158.1 MB · 16.0×
Claude Code~212.7 MB · 21.5×
OpenCode~318.4 MB · 32.2×
Next step
Choose a workflow before assembling a stack.
The smallest tool set that passes a real acceptance test is usually the easiest one to secure, explain and replace.