Are AI Coding Agents & Frameworks Actually Open Source?

<a href="https://theagentsindex.com/blog/open-source-license-census-ai-agents-2026/" target="_blank" rel="nofollow noopener"><img src="https://theagentsindex.com/charts/open-source-license-census-ai-agents-2026.png" alt="How the 33 open-source listings actually clear the bar — The Agents Index" title="How the 33 open-source listings actually clear the bar — The Agents Index" width="640" loading="lazy" style="max-width:100%;height:auto" /></a>

Of the 94 AI agent products published in The Agents Index on 27 August 2026, we now have a determined license for all 94. Thirty-three of those 94 are open source in some form, and 61 are not. License terms decide whether you can fork a tool, self-host it, resell it, or keep running it after the vendor is acquired or shuts down, so we read the license itself rather than the vendor’s adjective for it.

“Open source” shows up in the marketing copy of products that turn out to be proprietary SaaS with a free tier. It also runs the other way: some genuinely open, self-hostable frameworks barely use the phrase. The corpus also holds one agency listing, Agentix Labs, which sells engineering services rather than software with a license to read, so it sits outside this count entirely.

What changed since the last edition

The previous edition, verified on 25 August 2026, covered 93 listings with a determined license for every one of them. This edition adds a single new listing: Lead Scorer, an MCP-native AI SDR that joined Sales & marketing agents on 26 August 2026. It ships as a hosted MCP server and a web CRM with no public repository and no license to read, so it is closed source, landing in the 61 rather than the 33, and it moves neither the category split nor the open-core, caveated, or unread-license buckets. No listing left the index between editions, and no existing Yes or No changed.

That keeps the unknown bucket at zero: Lead Scorer arrived with a determined license rather than joining a queue, which is the only way this page stays current between editions. A month ago six listings had never been read; this edition, none do.

The headline count

Thirty-three of the 94 determined listings, or 35%, are open source in some form. That single number hides five quite different positions, and the difference matters if you are deciding what to build on.

Twenty-six carry a clean, unrestricted OSI-style license (MIT, Apache-2.0, GPL-3.0-or-later or AGPLv3) across their entire codebase, free to self-host with no carve-outs: LangChain, LangGraph, Microsoft AutoGen, Microsoft Agent Framework, LlamaIndex, OpenAI Agents SDK, Pydantic AI, CrewAI, AgentsKit, OpenCode, Zot and Kilo Code (all MIT); Cline, Aider, OpenAI Codex CLI, LiveKit Agents, Letta, GPT Researcher, Google Agent Development Kit, Cartesia, Agno and Goose (all Apache-2.0); Juggler (AGPLv3 app, Apache-2.0 extension SDK); Zed (GPL-3.0-or-later core, Apache-2.0 for select components); and Botpress and OpenClaw (both MIT), the only two agent platforms in the index with an unrestricted license.

Two ship an open core with a separately-licensed Enterprise layer. Mastra is clean Apache-2.0 at its core but gates its Enterprise controls (RBAC, SSO, IAM) behind a separate source-available license. OpenHands is MIT everywhere except its enterprise/ directory, which carries the PolyForm Free Trial License 1.0.0. In both cases the part you would actually self-host is fully open.

Two call themselves open source with a real license caveat. n8n is “fair-code”, a source-available license that is explicitly not OSI-approved; n8n’s own cons admit “check terms for commercial resale”. Dify ships an “Apache-2.0-derived” license with added conditions our sourced research flags as “not pure OSS”.

One open-sources the shell and not the engine. Warp’s terminal client went dual AGPL-3.0/MIT in April 2026, but Oz, the cloud platform that actually runs and coordinates its agents, was never part of that release and stays closed. That is a different caveat from Mastra and OpenHands: for Warp the part doing the agentic work is not open at all, only the interface around it.

And two are open source without a license we have read. Sim is counted as open source on the strength of its self-hosting option and a 29.4k-star public repository, both cited on its listing, but our own research does not name the license it ships under. Vellum is the same shape: its own listing states it is open source and self-hostable, without naming which license covers that code. Both are counted in the 33 and excluded from the 26, and this note exists so nobody mistakes one bucket for the other.

The remaining 61 determined listings are closed-source products. There is no code to inspect, no self-hosting option, and your continuity guarantee is the vendor staying in business. That is a live question: Moveworks went from independent company to ServiceNow subsidiary when its $2.85B acquisition closed in December 2025.

The 26 clean-license entries are also maintained as a standing collection you can bookmark, Open Source AI Agent Tools, and its 26 members are exactly the 26 above.

The category split is the real finding

Open-source rate is not spread evenly. It tracks almost exactly with who the buyer is: infrastructure a team assembles is usually open, and a finished business outcome almost never is.

Agent frameworks are 13 of 14 open source, and sales, marketing and support agents are 0 of 17 between them. That is the widest gap on this page and it is the number worth taking away from it.

Sierra, Botsify and ElevenLabs Conversational AI each belong to two categories and are counted once in each, so the category rows above cover 97 memberships across 94 distinct tools.

Open-source AI coding agents

Coding agents are the closest category to an even split: 11 of the 25 in the category are open source and self-hostable, and every coding agent in the index has a determined license. They are Cline, Aider, OpenAI Codex CLI, OpenCode, Zot, OpenHands, Juggler, Zed, Kilo Code, Warp’s terminal client, and Goose, which is Apache-2.0 and now maintained by the Agentic AI Foundation under the Linux Foundation rather than by a single vendor.

The other 14 are proprietary IDE or SaaS products: Cursor, GitHub Copilot, Claude Code, Devin, Windsurf, Replit Agent, Bolt.new, Factory, CodeRabbit, Antigravity CLI, Muse Code, Junie, Augment Code and Gemini Code Assist.

Gemini Code Assist is the one worth reading twice. Subscribers also get access to Gemini CLI, which is a genuinely Apache-2.0 Google project in its own repository, but that is a separate product; Code Assist itself, the IDE integration you are actually buying, ships nothing open.

Read the license note before assuming “open-source coding agent” tells you the whole story. Warp is the edge case in the open group: its terminal shell is genuinely open source, but Oz, the cloud platform that runs and coordinates its agents, is closed and cannot be self-hosted.

Open-source AI agent frameworks

Agent frameworks are 13 of 14, the highest rate in the index, and that is not a coincidence. A framework’s entire value proposition is code you import and run yourself, so a closed-source agent framework is usually a contradiction in terms.

The 13 are LangChain, LangGraph, CrewAI, Microsoft AutoGen, Microsoft Agent Framework, LlamaIndex, OpenAI Agents SDK, Pydantic AI, Mastra, Letta, Google Agent Development Kit, Agno and AgentsKit. Mastra is the one caveat among the 13, since its Enterprise controls sit behind a separate source-available license.

Claude Agent SDK is the sole framework marked No. Its Python and TypeScript wrapper packages carry a genuine MIT license, but the compiled Claude Code CLI binary they bundle and spawn is not published as source, and Anthropic’s own docs place overall SDK use under its Commercial Terms of Service. A thin shim around a closed engine is not an open product with a closed companion, which is why it counts No here and Cartesia’s Line SDK counts Yes: Line does substantial work of its own, including LLM routing, tool calling and multi-agent handoffs, and calls Cartesia’s proprietary speech models over the network the way an open CLI calls a closed model.

Agent platforms, and the three categories at zero

Agent platforms are 6 of the 16 where a license is determined, and every agent-platforms listing now has one. n8n and Dify are the two caveated cases above, Sim and Vellum are the two with no license named in our research, and Botpress and OpenClaw are the two with clean, unrestricted MIT licenses. The other ten, Lindy, Relevance AI, Salesforce Agentforce, Sierra, Moveworks, Glean, Botsify, Vecbase, MindStudio and ElevenLabs Conversational AI, are fully proprietary. MindStudio is the listing whose license was undetermined two editions ago; it is now confirmed closed.

OpenClaw is the more demanding of the two clean options to actually operate. Its own listing records a reviewer spending roughly six hours on a VPS install, with no task dashboard once it was running.

Sales and marketing agents are 0 of 8, support agents are 0 of 9, and agent tools and infrastructure is 0 of 2, with no undetermined listings in any of the three. These are the categories sold as a finished business outcome, leads booked or tickets resolved, rather than as infrastructure a team assembles. Buyers here are choosing a vendor relationship with no self-hosted fallback if that relationship ends.

The agent-tools category is the newest and the most counter-intuitive of the three zeros, because both of its members publish open-source code. Browserbase maintains Stagehand, a well-used automation framework anyone can run without an account, and SalesTouch’s public repository carries an MIT license. In both cases the open artefact sits beside the product rather than being it: Browserbase’s hosted browser infrastructure is closed, and SalesTouch’s repository holds a connector manifest pointing at a closed hosted endpoint rather than any application code. Both count No.

The full table

Every row below is read from the sourced research published on that tool’s own listing. The Yes/No column is generated from the same openSource field the rest of this page counts, so it cannot drift from the counts in the prose.

* The asterisk flags three different kinds of caveat. n8n and Dify both describe themselves as open source, and both are flagged because their own materials name a specific condition, fair-code terms in one case and an added-conditions Apache derivative in the other, that stops short of an OSI-approved license. Warp’s caveat is about scope rather than license purity: the AGPL-3.0/MIT license on its terminal client is clean, but it covers only the client. Sim’s and Vellum’s caveat is about our own evidence rather than about the product: the open-source claim is sourced, the license name is not. Read the note, not just the column, before assuming any of the five behaves like MIT in practice.

Why the caveats matter more than the yes/no column

A “No” in the table is unambiguous. You are a customer of a company, and your continuity depends on that company. The “Yes” rows are where the due diligence actually sits, because open source is not one guarantee.

A clean OSI license (MIT, Apache, AGPL) means you can fork it today and run it forever, and nothing the vendor does to its business changes that. This is the strongest continuity position on the page, and 26 of the 94 determined listings hold it.

Source-available or fair-code (n8n) and added-conditions (Dify) means you can read the code and usually self-host it free, but commercial resale, hosting-as-a-service or certain enterprise uses may be restricted. The vendor has retained rights an MIT or Apache license would have given away. Read the license text before you build a commercial offering on top.

Open core with a proprietary layer (Mastra, OpenHands) means the part you would run is open, but enterprise controls such as SSO, RBAC and hosted multi-tenancy sit behind a separate paid or source-available tier. That is a normal, disclosed pattern rather than a bait-and-switch, and worth knowing going in.

Open interface with a closed engine (Warp) goes a step further than open core. It is not that extra controls are paywalled; the component doing the actual agentic work is not released under any open license at all.

Sourced as open, license unread (Sim, Vellum) is a statement about the limits of this census rather than about either product. Go and read the repository yourself before you rely on the terms.

None of this is a knock on any vendor. Fair-code and open-core are legitimate business models. The point is that “open source” on a landing page answers a marketing question and the license file answers the buying question, and they are not always the same answer.

Methodology

This is a census of our own index, measured on 27 August 2026, not a survey of the market. It covers all 94 product listings published in The Agents Index on that date across eight categories, plus one agency listing held outside the count. We do not claim it describes the AI agent market; it describes what we have indexed and read.

Every row is read from the sourced, quality-gated research already published on each listing: the same openSource field and license language each listing cites to a primary source, usually the tool’s own GitHub repository, pricing page or licensing docs. A tool is marked open source only where its own listing states it with a citation.

The license name is held to a second, stricter bar: it is quoted only from that same sourced text, never from a vendor’s own description of itself. That is the entire reason Sim and Vellum are counted among the 33 open-source listings and not among the 26 clean-license ones.

Every listing now has a determined license. That has not always been true of this page, and when it was not, the undetermined listing was named in full rather than folded into either count.

Sierra, Botsify and ElevenLabs Conversational AI each belong to two categories and are counted once in each, so the category table covers 97 memberships across 94 distinct tools. The same convention is used by our deployment census. Both figures move as we add and re-verify listings, and each edition of this page is re-derived from the corpus rather than amended in place.

Three figures were checked a second way before publication. The 94-row table was diffed against the previous edition’s 93: Lead Scorer joined as a closed-source Sales & marketing agent, no listing left, and no other existing Yes or No changed. The 33 open-source listings were cross-checked against the open-source taxonomy tag. The 26 clean-license entries were checked against the membership of the Open Source AI Agent Tools collection, which lists the same 26 by name; that collection has not yet added Vellum to its own count as of this check, which is why this page states the discrepancy rather than silently matching a number that has not caught up.

For the vendor lock-in angle beyond licensing, covering ownership, funding and acquisition risk, see our AI agent acquisition wave field guide. For a narrower cut on six terminal coding agents, covering license and model portability together, see Terminal AI coding agents: open, portable, or locked in?. The Open Source Initiative’s license list is the reference we used to decide which licenses here are OSI-approved.