Which AI agent frameworks are actually open source — and what the licence covers in production

The short answer

Of the 18 agent frameworks we track, 15 ship a licence file, and 14 of those are MIT or Apache-2.0. So the licence is almost never what bites you. What changes when you go to production is one of four other things — and only one of them is written in the LICENSE.

Skip to the verdict if you want it now:

Everything below is verified against primary sources on 13 August 2026.

What the repos actually say today

The three we hold no licence data for — Agent.ai, Vertex AI Agent Builder and LangGraph Platform — are proprietary by construction. There's no repo to check.

One correction to our own records while we're here: our database had AutoGen down as CC-BY-4.0, which would alarm any lawyer who read it. The repo carries two files. LICENSE is Creative Commons and covers the documentation; LICENSE-CODE is a standard MIT grant from Microsoft Corporation and covers the code. AutoGen's code is MIT. We've flagged the record.

The four ways "open source" stops being true

1. The licence was never open source

Two popular tools in this space fail the OSI definition outright, and both are commonly described as open source anyway.

n8n ships under the Sustainable Use License, not an open source licence at all. The text is explicit: "You may use or modify the software only for your own internal business purposes or for non-commercial or personal use. You may distribute the software or provide it to others only if you do so free of charge for non-commercial purposes." Internal automation is fine. Wrapping it and selling it is not.

Dify uses Apache 2.0 with additions, and the additions are the point: "Unless explicitly authorized by Dify in writing, you may not use the Dify source code to operate a multi-tenant environment," plus a clause forbidding removal of Dify branding from the console. If your product is a workspace-per-customer SaaS, that clause is aimed directly at you.

Neither is a scandal — both are reasonable commercial choices, clearly published. But they are the only two here where the licence text itself restricts what you may ship, and that makes them the only two where legal review needs to happen before you write code, not after.

2. The repo is open, but it isn't the product

Botpress has 14,862 stars on an MIT-licensed repository. Read the README and you find it contains "Integrations… Devtools… Bots" — SDKs, a CLI, an API client and examples. The same README says outright: "This is not the recommended way to build bots and is in no way a replacement for the Botpress Studio." The Studio and the runtime live in Botpress Cloud.

There was a self-hostable Botpress. v12 was dual-licensed AGPLv3 and proprietary, and Botpress's own docs now state it "has been officially sunset and is no longer available for purchase, download, or new deployments," explicitly including "all other self-hosted or locally installed versions." That's the cleanest example in the category of a star count that measures a thing you cannot actually run.

If you're evaluating on GitHub metrics alone, this is the failure mode to watch for: check that the repo contains a runtime, not just a client.

3. The framework is free; the operations layer is the business

This is the normal case, and it's fine — as long as you price it in.

LangGraph is MIT and always will be. But LangGraph's own README points you to LangSmith Deployment for "long-running, stateful workflows," and LangChain's docs are blunt about what self-hosting that costs: "Self-hosted LangSmith is an add-on to the Enterprise plan designed for our largest, most security-conscious customers," with a licence key obtained by contacting sales. The published tiers stop at Plus, $39/seat/month, cloud-only. There is no self-serve path to running the platform in your own VPC. You can absolutely run bare LangGraph on your own infrastructure with your own observability — plenty of teams do — but you're then rebuilding the part LangChain sells.

CrewAI is MIT and says so, and its README is unusually straight about the split: "For organizations that need a commercial control plane around CrewAI, CrewAI AMP Suite adds managed deployment, observability, governance, security, and enterprise support." The free tier is 50 workflow executions a month — a demo allowance, not a pilot allowance. CrewAI also collects anonymous telemetry by default; the README documents what it excludes (prompts, task descriptions, agent backstories, tool usage, API calls, responses) and how to switch it off with OTEL_SDK_DISABLED=true. Disclosed and defeatable, but set it deliberately rather than discovering it in a security review.

Agno is Apache-2.0, verified in the licence file. The framework and a local Control Plane are free; the live Control Plane for production is $150/month, with $30/month per extra seat and $95/month per additional live connection. Self-hosting the Control Plane is Enterprise-only.

The pattern is identical in all three: permissive core, proprietary control plane, and the boundary sits exactly where "we built a demo" becomes "we operate this." Budget for the control plane or commit to building one.

4. The licence is fine and the project is leaving

AutoGen is the most-starred framework in our index at 60,393 stars, and its last push was 15 April 2026 — four months ago, in a field where every other maintained project here pushed this week. The README explains why: "AutoGen is now in maintenance mode. It will not receive new features or enhancements and is community managed going forward. New users should start with Microsoft Agent Framework."

Semantic Kernel, 28,443 stars and pushed two days ago, is going the same way. Its README opens with "Semantic Kernel is now Microsoft Agent Framework! Microsoft Agent Framework (MAF) is the enterprise-ready successor to Semantic Kernel."

So two of the highest-profile "open source agent frameworks" — both MIT, both permissively licensed forever — are converging on a single successor. Microsoft Agent Framework is itself MIT with 12,800 stars. Starting a new Microsoft-stack project on either predecessor today means writing a migration you already know about, and Microsoft's own guidance is that moving substantial multi-agent orchestration code is closer to a rewrite than a port.

Stars measure how many people liked something once. Last-commit date measures whether anyone is still there.

The governance question almost nobody checks

A permissive licence is irrevocable for code already published. It says nothing about the next release.

When one company owns the entire copyright, it can relicence future versions unilaterally — the move HashiCorp, Redis, Elastic and others have all made. Every framework in this table except one is controlled by a single company: CrewAI Inc, LangChain, Agno, deepset, Letta, Google, Microsoft, Botpress. Your protection is a fork of the last permissive commit, and a fork you have to staff.

goose is the exception. Block donated it to the Agentic AI Foundation, hosted by the Linux Foundation, and the repository moved from block/goose to aaif-goose/goose — which is why link-checkers and stale bookmarks have been failing. The same foundation anchors Anthropic's Model Context Protocol and OpenAI's AGENTS.md. Foundation stewardship doesn't guarantee good software, but it does mean no single vendor can change the terms on you later. At 52,744 stars, pushed today, Apache-2.0 with its explicit patent grant, goose is the strongest answer here if a licence change would genuinely hurt you.

Worth noting the split more generally: Apache-2.0 (goose, Agno, Haystack, Letta, ADK, CAMEL, Swarms) includes an express patent grant; MIT (LangGraph, CrewAI, PydanticAI, Semantic Kernel, Botpress, Dust, Agency Swarm) does not. For most teams that's academic. For anyone shipping into a patent-sensitive industry, it's the first thing your counsel will ask about.

So what should you actually use

Default for a Python team that wants to own its stack: PydanticAI. MIT, pushed today, 19,260 stars, and the paid product — Logfire — is observability you can decline and replace with any OpenTelemetry backend. Nothing in the runtime pushes you toward a vendor.

If your problem is retrieval-shaped rather than agent-shaped: Haystack. Apache-2.0, pushed today, and deepset's Enterprise Platform is genuinely optional rather than the place the real functionality lives.

If you want multi-agent structure and you're honest that you'll want a control plane: CrewAI for role-based crews, Agno if you want the runtime and API scaffolding included. Pick either, then budget the control plane cost — $150/mo for Agno's, sales-quoted for CrewAI's — from day one rather than discovering it at launch.

If you need durable state and memory as a first-class concern: Letta, Apache-2.0, with local and self-hosted server modes alongside its cloud.

If a vendor relicence would be an existential problem: goose, on governance grounds alone.

If you're deep in Microsoft: Microsoft Agent Framework, and skip both predecessors.

If you're building multi-tenant SaaS: rule out n8n and Dify at the shortlist stage. This is the one case in this whole piece where the licence, not the business model, is the blocker.

Before you commit, check these four things

1. Does the repo contain a runtime, or only a client? Botpress is the cautionary tale.
2. What's the last commit date, not the star count? AutoGen has 60,393 stars and four months of silence.
3. What does self-hosting the observability/deployment layer cost? For LangChain's stack, the honest answer is "call sales."
4. Who owns the copyright? One company means one signature is all it takes to change the terms on the next release.