SkillClaw: A Background Skill Evolution Loop for Hermes and Other Agents

What SkillClaw Solves for Single and Multiple Agents

SkillClaw addresses a specific pain: agents like Hermes accumulate skills from sessions, but nobody curates them. The README describes the result as duplicates, outdated entries, and half-baked skills piled together. SkillClaw's pitch is that it acts as a post-task digestion loop. It does not make the agent learn more; it makes what the agent already learned count. The target user is someone who runs one or more Hermes agents, possibly across devices, and wants a unified skill library. The project extends to teams: when several users share a group, each user's real-world experience feeds into the same evolution loop. User A debugs a database issue, the skill evolves, and users B, C, and D benefit without hitting the same problem. That collective angle is the core differentiator from a simple skill manager.

The Two-Loop Architecture: Task Time and Post-Task Evolution

The repository presents a two-loop diagram. The first loop is the agent's normal task-time operation: you chat, the agent executes, and it produces skills or learns from the interaction. The second loop is SkillClaw's contribution, running after the task. It takes the raw outputs from sessions and evolves them. The README uses the term 'Agentic Evolver' to describe this process. The evolution loop appears to be asynchronous and silent: you talk to your agent as usual, and skill evolution happens in the background. For multiple agents, SkillClaw merges, deduplicates, and cross-pollinates skills into a unified library, then distributes them back to all agents. The same mechanism works across devices, so a Home Hermes that learns React and a Work Hermes that learns Kubernetes share those skills. The architecture implies a central store or daemon that aggregates session data and writes back refined skills.

Installation and Commands: Setup, Start, and Dashboard

SkillClaw offers a shell installer for macOS and Linux, plus a manual Python install path for Windows. The README shows the commands `skillclaw setup` and `skillclaw start --daemon`. A terminal graphic in the README repeats those two commands, suggesting they are the primary workflow. There is also a dashboard command mentioned in the news section: `skillclaw dashboard sync`. The project requires Python 3.10 or higher. The README does not provide the exact shell command for the installer, nor the manual pip command for Windows. It also does not list configuration keys or environment variables. You would need to clone the repository or read the docs to get the full setup details. The presence of a daemon mode implies it runs persistently in the background, which is consistent with the 'silent evolution' promise.

Compatibility: Which Agents and APIs Are Supported

The badge list names Hermes, OpenClaw, Codex, Claude Code, QwenPaw, IronClaw, PicoClaw, ZeroClaw, NanoClaw, and NemoClaw. The README also says 'and any OpenAI-compatible API.' That is a broad claim. The actual integration depth likely varies: Hermes and OpenClaw have dedicated badges, while others are listed with a generic 'supported' badge. The mention of Codex and Claude Code is notable because those are closed-source commercial tools. How SkillClaw captures sessions from them is not explained in the README. It may rely on log files or API hooks. For an engineer evaluating this, the key question is whether your agent exposes session data in a format SkillClaw can read. The README does not specify that, so you must check the source code or documentation.

The Evolution Loop: What It Actually Does to Skills

The README uses terms like 'auto-evolve, auto-deduplicate, auto-improve quality.' It describes a skill evolution flow across multiple users with a diagram. The mechanism is not detailed in the README beyond the concept. There is an arXiv paper linked (arXiv:2604.08377), which presumably explains the algorithm. The project calls itself an 'Agentic Evolver,' implying that an LLM or agentic process performs the evolution, rather than a fixed rule-based deduplication. That is a meaningful design choice: it means the quality of evolution depends on the underlying model's capability. The README does not state which model is used or whether it is configurable. It also does not say how skills are represented (e.g., as text files, JSON, or a database). Without that, you cannot predict storage overhead or migration effort.

A Genuine Limitation: The 'Silent' Background Loop May Be Opaque

The biggest limitation is that skill evolution happens silently in the background. That is a feature, but it is also a risk. If the evolver makes a mistake, you may not notice until a skill degrades. The README does not describe any review or rollback mechanism. There is no mention of version control for skills, no audit log, no way to approve changes before they are distributed back to agents. The dashboard command suggests some visibility, but the README only mentions it in passing. For a team using shared skills, a bad evolution could propagate to all users. The project also does not state how it handles conflicting skills from different users or devices. The README says 'merged, deduplicated, and cross-pollinated,' but the conflict resolution policy is absent. That is a real gap for production use.

Alternative Approaches: Manual Skill Curation vs. Agent-Native Learning

The obvious alternative is to do what SkillClaw automates: manually curate your agent's skill library. Hermes and other agents already have a way to store skills, and you can edit them yourself. That gives you full control and no background daemon, but it does not scale across agents or users. Another alternative is to rely on the agent's built-in learning, if it has one. Some agents improve from feedback within a session, but that learning is often session-bound. SkillClaw's difference is that it operates across sessions, agents, and users, which is a fundamentally different scope. A third alternative is to use a vector database or a simple file sync tool to share skills across devices, but that does not perform evolution or deduplication. So the real comparison is between a centralized, automated evolution loop and a distributed, manual approach.

Maintenance and License: MIT, but Young and Rapidly Moving

The project is licensed under MIT, which is permissive and allows commercial use without copyleft obligations. That is a low barrier for adoption. However, the repository has no releases listed, and the last push was on 2026-08-17. The news section mentions recent activity, including a dashboard feature added on 2026-04-22. The lack of tagged releases means you are pinned to the main branch, which can change without notice. The README is heavy on marketing language and light on technical details, which is a red flag for an engineering audience. You must read the source to understand the data flow. The project also appears to be part of a larger ecosystem (AMAP-ML), with a related project LongHorizon-Harness. That suggests ongoing investment, but it also means the API may shift as the ecosystem evolves. Before adopting, you should verify that the project is actively maintained beyond the last push and that the paper matches the implementation.

Editorial conclusion

Adopt SkillClaw if you run Hermes or OpenClaw agents and want a shared, deduplicated skill library that improves from real use without manual curation. Skip it if you need offline operation, have no cross-session or cross-user needs, or cannot tolerate a daemon that watches every session. Before adopting, verify the claimed integrations with your specific agent version, test the setup and dashboard commands in a sandbox, and review the arXiv paper for the evolution algorithm's actual behavior.

Official sources

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