devtools
ranked
05 Sept 2026
Seven self-hosted AI agents, with the numbers pulled from the API not the blog posts
Every comparison article I found was stale by more than a hundred thousand stars. Here are the seven, ranked on what the repository itself says about them today.
By Purvansh Parmar
Checked 2026-09-05
Read 8 min
Entries 7
Two projects have taken over this category in about nine months, and almost everything written comparing them is wrong on the numbers. Not slightly wrong. One article I read understated a project by 209,000 stars, which is more stars than all but a handful of repositories on GitHub have in total.
So I stopped reading and started querying. Everything below came from teh GitHub API on one afternoon, and I have said which date so you can watch it go stale too. What the numbers show is less about popularity than about who is drowning.
How this was checkedI read six comparison articles, then queried the GitHub API for the same repositories on 5 September 2026. Not one article was close. OpenClaw was quoted at 250,000 and 350,000 stars by two different pieces; the API returned 388,913. Hermes Agent was quoted at 32,000 and at 140,000; the API returned 241,692. The ratio nobody mentioned was open issues: Hermes carries 39,901 against 241,692 stars, while OpenClaw carries 6,227 against a larger number.
What I compared on
- stars, licence and last push, read from the API rather than a blog post
- open issues measured against the size of the audience
- what it is actually for, since three of these are not competing
- what it costs to run once the software is free
A gateway that wires chat apps to agents running on hardware you own.
Wins
It is the broadest of the seven and the most obviously production-minded. Released v2026.9.1 on 3 September and pushed again the day I checked, which for a project this size means real staffing rather than a founder and a weekend.
Loses
The API does not resolve its licence to a standard identifier, returning NOASSERTION instead of a name. That is a genuine problem if anyone in your organisation has to approve dependencies, and it is the one thing no comparison article I read mentioned.
388,913 stars6,227 open issuesv2026.9.1NOASSERTION licence
A persistent agent that keeps memory between sessions and turns repeated work into reusable skills.
Wins
MIT licensed, which makes the adoption conversation short. Memory that survives a restart is a genuinely different product from a chat window, and the skill system means the second time you ask for something it does not start from nothing.
Loses
39,901 open issues against 241,692 stars. That is roughly one open issue for every six people who starred it, and six times the ratio OpenClaw carries at a larger size. Filing a bug there is closer to a message in a bottle than a support request.
241,692 stars39,901 open issuesMITv2026.8.31
Natural language in, code written and executed on your machine out.
Wins
Eleven open issues. Not eleven thousand. For a project with 68,241 stars that is the cleanest tracker in this list by two orders of magnitude, and it says the maintainers are closing things rather than accumulating them.
Loses
It is narrower than the two above and does not pretend otherwise. There is no chat-app gateway and no persistent memory, so if you want an assistant that remembers last week you are building that part yourself.
68,241 stars11 open issuesApache-2.0
An extensible agent that runs locally and drives your tools through MCP.
Wins
Pushed the day I checked, Apache licensed, and 300 open issues against 53,926 stars is a healthy ratio. It started inside a large company and has since moved to a foundation-style organisation, which usually means governance someone thought about.
Loses
That move is also the risk. The repository now lives at a different owner than most documentation points to, so half the links you find redirect, and a redirect is a thing that works until someone decides it should not.
53,926 stars300 open issuesApache-2.0
A personal assistant that indexes your own documents and answers from them.
Wins
The retrieval story is the point here, and it is the one of the seven that starts from your files rather than from a task. 147 open issues against 37,125 stars is comfortable.
Loses
AGPL-3.0. If you plan to run a modified copy as a service for anyone else, that licence reaches further than the others here and your legal team will want to talk about it. For personal use it costs you nothing.
37,125 stars147 open issuesAGPL-3.0
The memory layer itself, exposed as a service you build agents on top of.
Wins
39 open issues. It is the most focused project in the list and the only one that treats memory as the product rather than a feature, which is why it turns up underneath other people's agents rather than competing with them.
Loses
It is not an assistant and will not behave like one. If you came here wanting something to answer your Telegram messages, this is a library you would have to build that thing on, and the smallest star count here reflects a narrower audience rather than a worse project.
24,623 stars39 open issuesApache-2.0
Pair programming in the terminal, editing files in a git repository directly.
Wins
Still the most pleasant of these to use on an existing codebase, and 48,755 stars says a lot of people agree. Apache licensed, and the git integration remains the sharpest implementation of the idea.
Loses
Last pushed on 22 May, which was three and a half months before I checked, while 1,855 issues stayed open. In a category where two competitors shipped releases within the last week, a quiet quarter is a signal and not a coincidence.
48,755 stars1,855 open issueslast push 2026-05-22
Side by side
| Project | Stars | Open issues | Licence | Last push |
|---|
| OpenClaw | 388,913 | 6,227 | NOASSERTION | 2026-09-05 |
| Hermes Agent | 241,692 | 39,901 | MIT | 2026-09-05 |
| Open Interpreter | 68,241 | 11 | Apache-2.0 | 2026-08-20 |
| Goose | 53,926 | 300 | Apache-2.0 | 2026-09-05 |
| Aider | 48,755 | 1,855 | Apache-2.0 | 2026-05-22 |
| Khoj | 37,125 | 147 | AGPL-3.0 | 2026-08-02 |
| Letta | 24,623 | 39 | Apache-2.0 | 2026-08-23 |
Verdict
Hermes Agent, if you want the assistant. The MIT licence and memory that survives a restart are the two things that matter most here, and the issue backlog is a reason to expect no support rather than a reason to avoid the software. Take OpenClaw instead when someone has to approve the dependency and a licence the tooling cannot name is a blocker, which in most companies it is. Open Interpreter is the pick when you want the code-execution loop and nothing else. Do not start something new on Aider this month. And check these figures yourself, because mine will go stale too.
Sources