SkillClaw: a post-task evolution loop for Hermes and OpenClaw skill libraries

The problem SkillClaw targets: a skill library nobody digests

The README frames the problem in terms of digestion rather than acquisition. Its own words: "The problem isn't that Hermes doesn't learn enough. It's that nobody helps it digest." An agent that appends a new skill after every session accumulates duplicates, stale entries and half-baked ones. Nothing in the task-time loop removes them, because removal is not a task.

SkillClaw's answer is a second loop that runs after the task. The README calls the pair "Two Loops: Hermes task-time loop + SkillClaw post-task evolution loop." The first loop is the agent doing work and recording what it learned. The second loop takes that record, evolves it, and writes the result back into the library the first loop reads from.

The intended audience is narrow and specific. You need an existing agent whose skills live somewhere on disk, and you need that library to be large enough that duplication is a real cost. A single agent with a dozen hand-written skills gains little. The project's own framing is a user who has "been using Hermes for a while" and whose library now looks like "an unsorted loot box." If that is not your situation, the evolution loop has no input to work on.

How the evolution loop is wired: two loops, a shared library, and a distribution step

The architecture image in the repository is captioned "SkillClaw Architecture," and the README describes the loop rather than the internals. What can be stated from the repository layout is the shape of the system: a Python package named skillclaw that exposes a CLI, a separate evolve_server package that can be deployed centrally, a dashboard asset bundle inside the skillclaw package, and a client_env.example.sh file at the top level for client-side configuration.

pyproject.toml registers two console entry points. The first is skillclaw, mapping to skillclaw.cli:skillclaw. The second is skillclaw-evolve-server, mapping to evolve_server.__main__:main. That split matters: the client that sits next to your agent and the server that performs evolution are separately installable and separately deployable.

The data flow the README describes is: a session produces experience, that experience feeds the evolution loop, the loop merges and deduplicates skills, and the result is distributed back. In the multi-agent case the README says skills are "merged, deduplicated, and cross-pollinated into a unified library, then distributed back to all agents." In the multi-device case the same library is shared across machines. In the team case, the README's claim is that one member's debugging session evolves a skill that other members then receive.

Two details are worth flagging. First, the README says evolution happens "silently in the background," which means the loop is a daemon rather than a command you invoke per session. Second, the optional dependency groups in pyproject.toml tell you which parts of this are core and which are bolt-on: embedding-based retrieval, LLM-driven evolution, cloud sharing and server deployment are all extras, not base requirements.

Installing SkillClaw and running a first real session

The README offers a shell installer for macOS and Linux plus a manual Python install path for Windows. The terminal graphic at the top of the README shows the two commands that follow installation: skillclaw setup and skillclaw start --daemon. The base package requires Python 3.10 or later, per pyproject.toml.

A manual install from the repository is the path that gives you the most control over optional extras. The uv invocation is documented in pyproject.toml itself under [tool.uv]:

uv sync --extra all

That installs every extra group: embedding, evolve, sharing and server. The [project.optional-dependencies] table in pyproject.toml defines those groups, and the all group is written as skillclaw[embedding,evolve,sharing,server].

After installation, run setup once. This is the step that connects SkillClaw to your agent's skill location and writes client-side configuration. The repository ships client_env.example.sh as a template for that client environment.

skillclaw setup

Then start the background loop. The README's headline example is exactly this:

skillclaw start --daemon

What you should see is the daemon running alongside your agent. From that point the README's instruction is to "just talk to your agent as usual." Skill evolution is described as happening in the background with no extra interaction. The check that matters is whether skills in your library change after a session; the README does not document a dry-run mode, so the first verification is a diff of the skill files before and after.

If you also want the bilingual dashboard mentioned in the news section, the README lists skillclaw dashboard sync as the command that was added for it.

What SkillClaw does not do, and when it is the wrong tool

The README is a product narrative, not a reference manual. It does not document rollback of an evolved skill, and it does not describe how to review or reject a change before it lands. For a system whose entire job is rewriting files your agent depends on, that is a real gap. If your skills encode procedures where a wrong rewrite is expensive, an automatic evolution loop with no documented review step is a poor fit.

The second limitation is structural. Evolution quality depends on the model doing the evolving. pyproject.toml puts openai under the evolve extra, and the README's compatibility list ends with "any OpenAI-compatible API." So the loop is only as good as the endpoint you point it at. A weak model asked to merge two overlapping skills can produce a merged skill that is worse than either original, and nothing in the described flow catches that.

The third is operational. This is a daemon plus, for shared scenarios, a server. The evolve_server package and the server extra (boto3, openai, oss2, python-dotenv) exist because collective evolution across users needs a central component. The sharing extra names Alibaba Cloud OSS specifically. A single developer who just wants better local skills is paying for infrastructure they will not use.

Finally, the README's collective story assumes trust. When "every team member's real-world experience feeds into the same evolution loop," one member's session can change the skills everyone else runs. The README does not describe access control or provenance tracking.

How SkillClaw differs from a self-improving agent loop

The closest alternative is the agent's own built-in learning. Hermes, OpenClaw and the other supported Claw agents already accumulate skills during task execution; SkillClaw's README explicitly separates the two, calling them the task-time loop and the post-task evolution loop. The difference in approach is timing and scope. A built-in loop adds knowledge while the task runs and cannot see across sessions. SkillClaw runs after the fact, which is what lets it compare two skills written on different days and decide they are the same skill.

A second point of comparison is the research line the related searches point at. The README links an arXiv paper, and the search terms around this project include SkillRL, SkillOpt, Trace2Skill and CoEvoSkills. Those names appear in the search data, not in the repository, so the README gives no comparison to them and I will not invent one. What can be said is that SkillClaw ships as a CLI plus daemon plus optional server, whereas a method paper typically ships as training code. If you want a deployable component next to an existing agent, the packaging is the difference that matters.

A third alternative is doing nothing and cleaning the library by hand. That is genuinely better for small libraries. Manual curation catches semantic errors that a merge pass will happily combine. SkillClaw's advantage only appears once the library is large enough that nobody is reading it anymore.

Maintenance, upgrade cost and the MIT licence

The repository is not archived and the last push was on 2026-08-17. There are no retrieved releases, so pyproject.toml is the version signal: 0.4.0. That means upgrades arrive through the repository rather than through tagged releases, and you should expect to track main if you want fixes.

The dependency surface is modest at the base and grows with the extras. Core is click, pyyaml, fastapi, uvicorn, httpx, tiktoken and requests. The embedding extra pulls numpy and sentence-transformers, which is the heavy one. The evolve extra pulls openai. The sharing and server extras pull boto3 and oss2. If you enable all of them you are maintaining a Python service, a model dependency and a cloud SDK in one install.

Upgrade cost is mostly configuration drift. The client reads a client-side environment file, and the repository ships client_env.example.sh as the template. When that template changes, your local copy does not update itself, so a diff against the example is the cheap check after pulling. The evolve_server package is deployed separately and has its own requirements-server.txt, so client and server can drift apart.

The licence is MIT, stated in the repository's LICENSE file and in the README badge. MIT is permissive: it allows commercial use and modification with attribution and no warranty. That is the licence text, not advice about your situation. One thing to note is that the optional sharing path names Alibaba Cloud OSS, so the terms you actually operate under include your cloud provider's, not just MIT.

Editorial conclusion

Adopt SkillClaw if you already run Hermes, OpenClaw or another supported Claw agent and your skill library has grown into duplicates and half-finished entries, or if several agents or teammates should draw on one shared library. Skip it if you have no existing skill corpus to digest, or if you cannot commit to running the daemon and the evolve server, because the loop does nothing on its own. Before trusting it, verify three things in your own checkout: that skillclaw setup finds your agent's skill directory, that skillclaw start --daemon actually writes evolved skills back after a session, and whether the sharing extra is required for your multi-device case. The repository is MIT licensed and the last push was on 2026-08-17.

Frequently asked questions

Official sources

1. AMAP-ML/SkillClaw on GitHub
2. Issues
3. License: MIT
4. README