Ouroboros: the agent ships pytest as a runtime dependency and needs Git to launch
Ouroboros, self-creating AI agent. Born Feb 16, 2026. Think between requests. Background consciousness supports reflection, initiative, and preparation outside the immediate request-response loop.
At a glance
- What is it?
- A self-modifying coding agent in Python that ships as seven platform binaries, keeps its own memory on your machine, and wraps a review apparatus around the one thing it advertises. The dependency comments and the Dockerfile explain more about how it is built than the feature claims do.
- Who is it for?
- Ouroboros suits someone who wants an agent that keeps identity and memory across tasks and restarts, and who is willing to read the contributing guide before letting it near its own source. Four things to weigh.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository received new commits within the last day.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- razzant/ouroboros1,669 ★ · +934
| Repository | 2026-07-30 | 2026-11-15 |
|---|---|---|
| razzant/ouroboros | 735 | 1.7k |
934 Stars in the last 108 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
Seven download buttons, all pointing at one tag
There are seven platform targets: a DMG for macOS 12 on Apple silicon, a ZIP for Windows x64, a deb for Debian, Ubuntu and Astra Linux, a generic RPM for Fedora and RHEL, a separate red80 RPM for RED OS 8, and an AppImage plus tar.gz for anything else on x86_64. Android exists as an experimental path that requires a Magisk-rooted ARM64 device, and its USB setup needs Python and adb, so it is not a peer of the other seven. The install page is explicit that you do not need to clone the repository or install Python or uv. One detail to check before you click: every one of those link targets is written against a single version, so the buttons do not follow new releases by themselves and will keep serving that version until the file is edited.
Checksums, an SBOM and smoke files that are not installers
The release carries its own verification trail, and the download section pre-empts a specific confusion about it. Four kinds of file sit alongside the binaries: a SHA256SUMS file, a release-evidence.json, per-run smoke files named release-smoke with a JSON suffix, and SBOM files with the CycloneDX naming. The documentation states plainly that these are verification evidence and not additional installers, which is a useful thing to say because a directory of JSON files next to a DMG invites exactly that error. Taken together they cover three separate questions: whether the download matches what was published, what the release process recorded when it ran, and what the artifact contains. A smoke file per run is the unusual one, since it implies the build executes and records an outcome rather than only hashing.
Git is a runtime dependency, because it rewrites itself
On the AppImage path the instructions say Git must already be installed, which reads like boilerplate until you notice the container image installs it deliberately, with git named as the dependency the agent needs for its own history and updates. That is the architecture showing through. The project advertises that it can rewrite the implementation it runs on, including its code, architecture, prompts, tools and dependencies, and it can also change how it understands itself without losing continuity. A process that edits its own source needs somewhere to record what it changed, and version control is that place, so Git is a functional dependency of the product rather than a convenience for developers. The same logic explains a first-run wizard for model access, review policy and budget: there is a control surface because the thing being controlled is the program itself.
Reviewed core evolution, with a lint gate that only checks one rule class
The self-modifying claim is deliberately paired with a review apparatus. The documentation opens with a warning that anyone changing the project, coding agents included, has to read the contributing guide first, and that the guide defines required project context, verification and a separate-agent review flow. A BIBLE.md sits at the root alongside the prompts directory, and the technical report is titled around reviewed core evolution and describes a 161-day deployment. The build side is narrower than you would expect. The lint target runs ruff with only the F rules selected, which the Makefile calls a deterministic gate for the NameError class, matching the quick-test step in CI. So style is not enforced, undefined names are, and that line is the one that keeps a self-editing codebase from rotting.
Why pytest and pip are runtime dependencies
Two entries in the dependency list look wrong until you read their comments. The first is pytest, flagged in the manifest as runtime rather than test-only, because the hermetic commit gate's parallel pass hard-blocks with a specific preflight-plugin-missing error instead of degrading to a serial run. A wheel that has to run that gate therefore has to carry the test framework. The second is pip itself, declared because the runtime self-update and install paths execute the interpreter's pip module, and uv-created environments do not seed pip unless it is asked for. A third file completes the picture: requirements.txt is now only a compatibility pointer for managed updaters released before the move to uv, and it delegates to a separate runtime lock file. Two dependency paths coexist, one legacy and one current.
Browsers are kept out of the package tree on purpose
The container image comment explains its own layer ordering with unusual candour. Browsers are installed first, dependencies second, sources last, and the reason is that every release rewrites the manifest and the lock file, so anything below the lock copy is rebuilt per release while the apt packages and the Chromium and WebKit downloads above it are reused. Playwright is pinned by argument at 1.62.0, and the comment insists the pin must equal the locked version so the downloaded browser revisions match the driver in the virtual environment, with a named test enforcing it. It also notes the installer runs from an ephemeral tool environment, so the image carries exactly one Playwright, the virtual environment's. Browsers then live at a shared path the runtime honours as-is rather than inside the package tree, because the package tree is what gets rebuilt every release. Running it is a single port mapping:
docker run --rm -p 8765:8765 ouroboros-webFrom v4.1 to v6.2.0 in forty-eight hours, thirty-two cycles
The project's own record of its start is specific enough to check. It first booted on February 16, 2026, and during the following forty-eight hours the repository advanced from the v4.1 line to v6.2.0. A self-authored record preserved from that period counts thirty-two evolution cycles. The first generation ran in Google Colab through Telegram, and it is still preserved on a branch named legacy-google-colab with its own archive page, while the current generation carries the same stated identity into a native desktop and headless runtime. Today the newest tag is v7.5.1 and the last push is dated 2026-10-02. The version numbers are worth reading as a trace of how fast this codebase moves: two minor versions in a day at the start, and a patch line in the seven range now.
The benchmark charts call themselves self-reported
Rather than quoting scores, the documentation is more useful about how they were obtained. The charts cover Terminal-Bench 2.1, OSWorld-Verified and CL-Bench, measured against Codex, Claude Code, Cursor and Hermes. The comparability caveat is stated in the same sentence as the claim: where a matched pair was run, the comparison was on the same model, and where it was not, the comparison is against the public leaderboard instead. That distinction matters, because a leaderboard number and a same-model rerun are not the same measurement, and saying which is which is rarer than it should be. The surrounding claim is that the agent can rewrite its own code and that this is reviewed rather than silent, which is the part a reader evaluating the project has to take on trust or go check.
Editorial conclusion
Ouroboros suits someone who wants an agent that keeps identity and memory across tasks and restarts, and who is willing to read the contributing guide before letting it near its own source. Four things to weigh. The download links are pinned to a specific tag rather than tracking the newest release, so check what you actually get. Git is not optional, because the agent's self-modification is built on its own history. pytest and pip are both shipped as runtime dependencies, which is a cost paid to keep a self-update path and a commit gate working. And the benchmark charts describe themselves as self-reported, with an explicit note about which comparisons were matched and which came off a public leaderboard.
Frequently asked questions
What is Ouroboros and what does it do between requests?
It is an open-source agent whose identity, durable memory and history continue across tasks and restarts. It works on external projects, coordinates a live swarm of specialist agents, and can rewrite the implementation it runs on, including its code, architecture, prompts, tools and dependencies.
Do I need Python or uv to install Ouroboros?
No. There are downloads for macOS on Apple silicon, Windows x64, a deb, an RPM, a RED OS package, an AppImage and a tar.gz archive. The optional CLI is installed from inside the desktop package, creates a user-local command without sudo, and also needs no Python or uv.
Why does Ouroboros require Git?
The container image installs git specifically for the agent's own history and updates, and the AppImage instructions say Git must already be installed. Because the agent can modify its own implementation, version control is part of how it operates rather than a developer convenience.
How does Ouroboros verify its releases?
Alongside the binaries it ships a SHA256SUMS file, a release-evidence.json, per-run release-smoke JSON files and CycloneDX SBOM files. The documentation states these are verification evidence and not additional installers.
Are the Ouroboros benchmark numbers official?
They describe themselves as self-reported results on Terminal-Bench 2.1, OSWorld-Verified and CL-Bench, against Codex, Claude Code, Cursor and Hermes. Comparisons use the same model where a matched pair was run, and the public leaderboard where one was not.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/razzant-ouroboros)Star history chart for your README
[](https://hysenlabs.com/projects/razzant-ouroboros)Was this useful?