The Best Open-Source AI Agent Projects on GitHub, Mapped

The best open-source AI agent projects on GitHub are no longer a single list — they are eight distinct lanes with different leaders, different licences and wildly different maintenance pace. This map covers 34 projects: coding harnesses and CLIs, agent frameworks, browser and computer use, memory, voice, sandboxes, observability, the neutral protocol layer underneath all of them, and the adjacent workflow and deep-research projects that sit beside them.

Every number below came from opening the repository itself — the star count from the repo page, the licence from the LICENSE file and GitHub’s own licence detector rather than a project’s marketing copy, and the activity signal from the last push. That extra step matters more than it sounds, because three of the loudest projects in this category are not open source under any standard definition, and one of them ships no CLI source code in its public repository at all.

What follows is the umbrella map: what each lane is for, who leads it, which projects are quietly abandoned despite still appearing on older best-of lists, and where the licence you assumed is not the licence you get. Deep dives on individual lanes are linked in place. Star counts and commit dates are as at the time of writing, and they move fast.

“Open source” has become a synonym for “the code is on GitHub,” and those are not the same claim. A public repository tells you nothing about whether you can fork it, modify it, self-host it commercially or read the part that actually runs. Three projects in this survey are routinely described as open source in coverage and fail that bar in three different ways — and all three are genuinely useful tools, which is precisely why the distinction gets skipped.

The clearest case is Anthropic’s Claude Code repository. At 140,531 stars it is one of the most-starred agent projects anywhere, and its LICENSE.md reads, in full: © Anthropic PBC. All rights reserved. Use is subject to Anthropic’s Commercial Terms of Service. GitHub’s licence detector returns no recognised open-source licence for the repo. The file tree holds plugins, scripts, examples, documentation and the changelog — the CLI itself is distributed as a binary through an install script, a Homebrew cask or npm. It is a public repository and a source-available distribution model, not an open-source project.

The distribution is the interesting part. Those three non-OSI projects are 8.8% of the 34 surveyed — and they hold 351,111 of the 1,730,761 combined stars, which is 20.3% of the total. In other words, one star in five in this category sits on a repository whose licence would not survive an OSI review. If your procurement process treats “has a GitHub repo” as a proxy for “open source,” it is wrong about a fifth of the market by attention.

None of this makes those projects bad choices. n8n is an excellent workflow engine and the Sustainable Use License is generous for the self-hosting case most teams actually have — we cover that ecosystem in depth in n8n’s AI node ecosystem. The point is narrower: read the LICENSE file before you build a business on the assumption, because the three failure modes above bite at three different moments — Claude Code when you want to audit or fork, n8n when you want to resell, Phoenix when you want to host.

Coding harnesses are the terminal agents that read your repository, plan an edit and run the commands. It is the single largest lane in this survey at 629,256 combined stars — almost three times the 213,257 of the framework lane — and it is where the licence gap is widest.

opencode leads on raw count at 194,461 stars, 53,930 ahead of Claude Code — roughly 1.38 times its total — and it does so with a genuine MIT licence and the actual CLI and TUI code in the repository. It is maintained by Anomaly, the organisation formerly operating as SST. Most “best open-source coding agent” round-ups still open with a vendor CLI, which is a ranking by brand recognition rather than by either stars or licence.

Split the same lane by licence and the picture sharpens: three of the four leading CLIs are open source, but the one that is not carries 140,531 of the lane’s 545,921 top-four stars — 25.7% of the attention. That is the shape of this whole category in miniature. The vendor harnesses win distribution; the community harness wins the licence. OpenHands is the interesting hybrid — an MIT-licensed end-to-end platform bundling the agent with its sandboxed runtime, rather than a bare CLI. Its GitHub organisation moved from All-Hands-AI to OpenHands, so older links and API paths against the previous org no longer resolve.

Frameworks are the libraries you reach for when you are writing the agent rather than driving one. The lane totals 213,257 stars across five projects, and its headline finding is not about capability at all — it is that the highest-starred framework here is also the least actively maintained.

The outlier is Microsoft AutoGen, the highest-starred framework in this survey at 60,284 — and the only one whose last push, 2026-04-15, is measured in months rather than days at the time of writing. Its licensing is also the most frequently flattened: GitHub’s detected primary licence is CC-BY-4.0 while the repository also surfaces an MIT notice, consistent with real dual-licensing where code is MIT and non-code content such as samples and documentation is CC-BY-4.0. Coverage that reports a single licence for AutoGen is reporting half of it.

Rounding out the lane, the OpenAI Agents SDK sits at 28,460 stars under MIT — close enough to smolagents that the choice between them is about ergonomics and provider affinity rather than momentum. For the head-to-head that this map deliberately does not re-litigate, see our comparison of the frameworks lane in more depth.

This lane is about agents that operate an interface rather than an API — clicking through a web app, filling a form, driving a desktop. It totals 181,955 stars, and browser-use alone holds 108,128 of them — 59.4% of the lane, and the fourth-highest count in the entire survey behind only n8n, opencode and Claude Code. It out-stars every agent framework in the previous section by a wide margin, which is a useful signal about where developer attention actually goes: not to the orchestration abstraction, but to the thing that touches the messy interface.

The project to strike off your shortlist is LaVague. It is Apache-2.0 with 6,385 stars and still surfaces on older best-of lists, but its last push was 2025-01-21 — more than eighteen months before this post. Listing it as a live option is the single most common error in this lane. The OSS-versus-commercial question is a separate one: for the vendor side, see our commercial computer-use matrix, and for the marketing-operations use cases these agents actually get pointed at, our guide to browser agents in marketing operations. The project-by-project comparison of this lane is this week’s sibling on open-source browser and computer-use agents.

Memory is the least standardised layer in the agent stack, and its star distribution reflects that: three serious projects, 91,683 combined stars, and Mem0 holding 68.4% of them. Mem0’s 62,736 stars also put it ahead of every framework in this survey — including AutoGen at 60,284 — which is a genuinely surprising result for a layer most agent tutorials still treat as optional.

Which of the three fits depends entirely on what “remembering” means for your application — extraction, graph or session — and that architectural comparison is this week’s dedicated sibling on Mem0, Letta and Zep compared. On the map, the useful fact is simply that all three are Apache-2.0 and all three are alive.

Voice tells the opposite story about where value concentrates. The lane totals 189,913 stars, but 159,442 of them — 84.0% — sit on the two model repositories rather than on any orchestration framework. The agent frameworks that turn a speech model into a conversation share the remaining 16%. If you are building a voice agent, the model layer is commoditised and well-starred; the pipeline layer is where you are still choosing between young projects.

whisper.cpp at 52,640 stars is the quiet load-bearing project here — the C/C++ port that underpins most local and offline dictation tooling, MIT-licensed and pushed to within days at the time of writing. Above it, Pipecat and LiveKit Agents are the higher-level pipeline frameworks and Vocode is the more modular, lower-level toolkit; Vocode’s repository has since been renamed to vocode-core, which older links miss. For the transcription pipeline side of this, see our guide to self-hosted Whisper transcription; for live dictation as a daily driver, this week’s sibling on open-source voice dictation versus the paid tools.

Sandboxes and observability are the operational lanes: where the agent’s code actually executes, and how you find out what it did. They are also the smallest lanes by stars — 13,282 and 49,635 respectively — which says more about adoption sequencing than about importance. Teams reach for a framework first and a trace viewer after the first incident.

The full end-to-end option in this space is OpenHands, counted in the harness lane above because it bundles the agent with the sandboxed runtime rather than exposing a sandbox primitive. If you already run a harness and just need isolation, E2B is the smaller commitment. If you need the whole loop under one MIT licence, OpenHands is the larger one.

The observability split is the practically important one. Langfuse and Helicone are permissively licensed; Phoenix is not, and the difference only surfaces at the moment you want to offer a hosted service on top of it. That is a narrow scenario — but it is exactly the scenario an agency or platform team ends up in, which is why the licence column belongs in the shortlist spreadsheet rather than in a footnote.

Underneath every lane above sits a protocol layer that no single vendor owns. The Agentic AI Foundation was formed under the Linux Foundation with three founding project contributions: Anthropic’s Model Context Protocol, Block’s goose, and OpenAI’s AGENTS.md, with Anthropic, OpenAI and Block anchoring it. Its platinum tier lists AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI; the gold tier includes Cisco, Datadog, Docker, IBM, JetBrains, Oracle, Salesforce, SAP, Shopify, Snowflake and Twilio; silver includes Hugging Face, Uber, WorkOS and Zapier.

That roster is why this layer belongs on the map. Most of the coding harnesses and frameworks surveyed above implement MCP, which makes it the one interface that survives you changing your mind about everything else in the stack. It is also a reminder that the interesting governance question in open-source agents is no longer “which licence” but “who decides the protocol.”

The star counts in this lane are a good lesson in reading repository metrics in context. Block’s goose carries 52,492 stars under Apache-2.0, while the MCP specification repository sits at 8,877 under MIT — roughly one-sixth of goose, and 5.9 times smaller. That is exactly what you would expect: goose is a tool people install, the spec repo is a document people implement. Judging a protocol by its spec repo’s stars would be like judging HTTP by the traffic to the RFC.

Speaking to the same announcement, Block’s head of open source, Manik Surtani, framed the moment as a decision point for a technology that will define the next decade — whether it stays closed or is built in the open. Anthropic’s Mike Krieger described MCP as having started as an internal project to solve a problem their own teams were facing, which is the honest origin story of most protocols that end up mattering.

Here is the whole map in one place: 34 projects, grouped by lane, with the star count, the licence as written in the repository, the activity signal and a one-line statement of what each is actually for. Last-commit cells read “Active” where the repository was pushed to within days at the time of writing; explicit dates appear only where the gap is large enough to change your decision.

Read the lane totals as an attention map rather than a quality ranking and a pattern falls out. Developers star the layer they touch: the harness they type into, the browser library that solves the thing APIs cannot, the model they download. They star the layers they operate — sandboxes and tracing — an order of magnitude less, even though those are the layers that determine whether an agent survives contact with production. The 47-to-1 gap between the harness lane and the sandbox lane is not a statement about engineering difficulty; it is a statement about when in the adoption curve each problem becomes visible.

Projected forward, that gap is the one most likely to close. The harness lane is already crowded with four near-equivalent leaders, which is what commoditisation looks like just before consolidation. The operational lanes are where a small number of projects still have room to define the category — and where the licensing pressure will land hardest, because tracing and execution are the layers that are easiest to turn into a hosted business. Expect more source-available relicensing there than in the harness lane, and expect the protocol layer to keep absorbing what everyone agrees on.

A map is only useful if it shortens a decision. Four come up repeatedly when teams sit down with a list like this one.

For marketing and operations teams weighing self-hosting against the SaaS tier, the honest cost is not the licence — it is the engineer who keeps the thing running; this week’s companion piece on open-source AI agents for marketing teams works through that total cost properly. If your question is specifically about research automation, the two projects in the adjacent lane above are compared in our sibling on open-source deep research agents. And if the decision is which of these to standardise on across a business rather than a single team, that is the shape of work our AI transformation engagements start with — an inventory, a licence audit, and one benchmark on your own workload rather than a leaderboard.