AI Agent Tools You Can Self-Host
28 of the 66 tools in our index can run somewhere you control — 11 with no managed-cloud alternative at all, and 17 that pair a self-hostable core with an optional paid managed layer on top. This is a deployment question, not a license one: 10 of the 28 are closed-source or license-hedged products with a genuine private-deployment path anyway.
Every listing in our index carries a `deployment` field — Cloud, Self-host, or Both — researched and sourced the same way as pricing or license. 28 of the 66 current tools set it to Self-host or Both, meaning code and data can stay inside infrastructure you control instead of a vendor's cloud, at some price. See our full [deployment census](/blog/ai-agent-deployment-census-2026/) for the category-by-category breakdown this collection is drawn from.
This list is deliberately not the same set as our [open-source collection](/collections/open-source-ai-agent-tools/), even though the two overlap heavily. All 18 tools on that list are also here, because a clean open-source license almost always means you can run the code yourself. But 10 more tools clear this list's bar without clearing that one: Mastra and Claude Agent SDK are open at the core but carry a license condition or commercial-terms wrapper that keeps them off the clean-license list; n8n and Dify are source-available/fair-code rather than a no-strings OSI license; and Factory, CodeRabbit, Bland AI, Windsurf, Warp and OpenHands are closed-source products that still sell (or in Windsurf's case, ship free) a genuine private-deployment or local-execution path. If a compliance team's real question is "can this run inside our own network," the license badge alone answers it wrong for a third of this list.
The category split is uneven in a way worth knowing before you shop by category page alone: agent frameworks are unanimous (11 of 11 offer some self-host path — a framework you cannot run yourself is arguably not a framework), coding agents are the most genuinely mixed category (12 of 19), and support agents plus sales/marketing agents offer zero — every one of those 14 tools is cloud-only at any price, so a hard data-residency requirement in either category currently has no architectural answer on this site, only a contractual one (a vendor DPA or regional hosting commitment).
Why this collection
A tool earns a place here if its own Published, sourced listing sets its `deployment` field to Self-host or Both, verified against the vendor's own docs or pricing page describing an actual self-run, on-premise, air-gapped, or local-execution option — not a roadmap promise or a generic "enterprise-ready" claim. This is a deployment-location question, independent of license: an entry may be open-source with no managed layer at all, open-source with an optional paid cloud add-on, source-available under a fair-code or added-conditions license, or fully closed-source with a paid tier that still lets you keep data inside your own infrastructure. For the license-only question, see the [open-source collection](/collections/open-source-ai-agent-tools/); for the full category-by-category count, see the [deployment census](/blog/ai-agent-deployment-census-2026/) this collection is built from.
- LangChain
The most widely-used framework for building LLM apps and agents, with LangGraph for stateful control.
No managed LangChain cloud product exists — MIT-licensed and self-hosted is the only way to run it, alongside the largest integration ecosystem of any framework here.
- CrewAI
A popular open-source Python framework for role-based, collaborating multi-agent teams, free to self-host, now a stable 1.0 release with an optional managed AMP cloud layer.
MIT-licensed with no managed CrewAI cloud offering — the orchestration layer runs entirely on your own infrastructure and model keys.
- Microsoft AutoGen
Microsoft’s open-source framework for orchestrating conversations between multiple AI agents.
MIT-licensed research framework with no managed layer — Microsoft's own successor, Agent Framework, is the one in this category that also offers a hosted option.
- OpenAI Agents SDK
OpenAI’s lightweight open-source framework for building agents, handoffs, guardrails and tracing in a few primitives.
MIT-licensed and self-host only — OpenAI does not offer a managed runtime for it, unlike Google ADK's Agent Runtime or Microsoft Agent Framework's Foundry layer.
- Pydantic AI
The type-safe agent framework from the Pydantic team, that FastAPI feeling for building AI agents.
MIT-licensed, Python-native, with no managed hosting layer from the Pydantic team — runs entirely inside a service you own.
- Cline
The open-source coding agent, an autonomous, bring-your-own-key agent that lives in your editor, terminal or SDK.
Apache-2.0 licensed with no cloud product to opt into — everything runs inside your own VS Code, JetBrains or terminal session.
- Aider
AI pair programming in your terminal, an open-source CLI agent that edits your code and commits to git.
Apache-2.0, a local terminal process only — there is no Aider cloud tier at any price, so running it yourself is the only way it exists.
- Juggler
An open-source, GUI-based workbench for AI coding agents, branching session trees, not a linear transcript. Free, AGPLv3, by JUCE creator Julian Storer.
AGPLv3 core with an Apache-2.0 extension SDK and no managed cloud version — the strongest copyleft license in this collection.
- OpenCode
A free, open-source (MIT) AI coding agent for the terminal, desktop and IDE, bring any of 75+ model providers, including your own local Ollama models.
MIT-licensed terminal agent with 188,976+ GitHub stars and no managed cloud counterpart — bring your own model key or a local model.
- Zot
A free, MIT-licensed coding-agent harness in one Go binary, branch any session from any past message and view it as a tree, plus a Telegram bridge.
MIT-licensed, a single Go binary with no server component at all — the lightest self-hosting story in this collection.
- GPT Researcher
An open-source autonomous research agent that plans, searches many sources and writes a cited report.
Apache-2.0 licensed — the only self-hostable pick in the research-agents category; every other listed research agent here is cloud-only.
- Microsoft Agent Framework
Microsoft’s production-ready, open-source successor to AutoGen and Semantic Kernel — one agent SDK for .NET, Python and Go.
MIT-licensed across four language SDKs — self-host free, or use Microsoft's optional Foundry Hosted Agents managed layer on top.
- LlamaIndex
The data framework for LLM apps, RAG-first, now with agent workflows and a document-parsing cloud.
MIT-licensed core self-hosts free; the paid LlamaCloud layer adds managed parsing on top, entirely optional.
- Mastra
The TypeScript-native agent framework — agents, durable workflows, memory and observability in one Apache-2.0 package, with an optional managed cloud platform.
Apache-2.0 core self-hosts free; a separate, optional Mastra Cloud platform adds hosting, a visual Studio and observability on a metered basis — open at the core, but that managed layer is why it sits outside the clean-license open-source collection.
- Letta
A stateful-agent framework — open-source, memory-first agents whose context, tools and identity persist as a durable server object instead of resetting between sessions.
Apache-2.0 licensed with persistent memory — self-host free, or opt into managed Letta Cloud later without changing your integration code.
- Google Agent Development Kit
Google's open-source, code-first agent SDK — four language runtimes, native A2A + MCP protocol support, and an optional managed Agent Runtime on Google Cloud.
Apache-2.0 licensed across all four of its language SDKs — self-host anywhere, or use Google's own Agent Runtime managed layer on top.
- Claude Agent SDK
Anthropic's own agent-building library — the same agent loop, tools and context management that power Claude Code, for Python and TypeScript. Claude-only, with no multi-provider escape hatch.
The wrapper package is MIT-licensed and the subprocess it spawns runs locally under your own control, but Anthropic's Commercial Terms of Service govern the SDK as a whole and the Claude Code binary it bundles isn't published as source — self-hostable in practice, just not on the clean-license open-source collection.
- OpenAI Codex CLI
OpenAI's open-source (Apache-2.0) terminal coding agent, bundled into ChatGPT plans from Free through Pro — Business seats are now grandfathered-only.
Apache-2.0 licensed CLI, free to self-host — OpenAI also offers a `codex cloud` managed alternative if you'd rather not run it yourself.
- OpenHands
Open source AI agent platform for cloud coding agents, from free local use to enterprise self-hosted deployment.
A free local OSS release (81,000+ GitHub stars) or a fully self-hosted enterprise deployment inside your own VPC — the hosted Cloud tier is the third, optional path, not the only one.
- Factory
Agent-native software development — specialized "Droid" agents that triage, code, test, review, document and ship changes across the SDLC, from a $20/mo CLI to a fully managed enterprise cloud.
The Enterprise tier spans fully managed SaaS, a hybrid cloud/customer-compute split, full on-premise, or air-gapped deployment across AWS, Azure or GCP with customer-managed encryption keys — the deepest self-hosting spread of any closed-source listing here.
- CodeRabbit
An AI code-review agent, not a code-writing one — reviews every pull request automatically across GitHub, GitLab, Bitbucket and Azure DevOps, with a genuine free tier for unlimited repos.
The Enterprise tier adds self-hosting and API access on top of the default managed SaaS code-review product.
- Warp
The AI-native terminal that replaces your shell and orchestrates Claude Code, Codex and its own Warp Agent side by side.
Open-sourced its terminal client under a dual AGPL-3.0/MIT license in April 2026, but Oz — the cloud platform that actually runs and coordinates agent work — stays closed and cloud-only; the self-hostable part is the terminal shell, not the agent backend.
- Zed
The open-source, GPU-native code editor built from scratch by Atom's creators — runs Claude Code, Codex and other agents via the open Agent Client Protocol it created. Free forever; Pro from $10/month.
GPL-3.0-or-later (Apache-2.0 for select components) — a full IDE you can self-host and modify, not just a CLI wrapper.
- Windsurf (now Devin Desktop)
The agentic IDE that was Codeium, then Windsurf, now Cognition’s Devin Desktop, a command centre for coding agents.
Devin Local lets the desktop IDE run its coding agent entirely on your machine instead of dispatching to Devin Cloud — a local-execution mode, not a team-deployable server, so weigh it differently than the enterprise on-prem entries here.
- n8n
Source-available workflow automation with native AI-agent nodes, self-hostable and integration-rich.
Source-available (fair-code) — self-host free with full functionality, or use n8n Cloud from €20/month; the license carries a condition worth reading before assuming it behaves like a no-strings open-source grant.
- Dify
Open-source, visual platform for building production LLM apps and agents, self-host it or run it in the cloud.
Apache-2.0-derived core (149K+ GitHub stars) — self-host with Docker for free, or use Dify's managed cloud; the built-in RAG pipeline runs the same either way.
- LiveKit Agents
The open-source, self-hostable voice agent framework, Apache 2.0, provider-agnostic, and the infrastructure behind ChatGPT’s Advanced Voice Mode.
Apache-2.0 licensed — the one clean open-source, self-hostable pick in an otherwise all-cloud voice-agents category.
- Bland AI
Enterprise voice AI that runs on its own infrastructure, build calls as conversational pathways, priced all-in per minute.
The Enterprise tier adds dedicated infrastructure with self-hosted or on-premises deployment, US/EU/APAC data residency, and SOC 2/HIPAA/PCI DSS compliance — the only voice agent here with a private-deployment path; every other listed voice agent is cloud-only.