English · 简体中文  |  🌐 omnilimb.com

Omnilimb

Your Hermes agent is the brain. Omnilimb is the hands and feet.

Built for the latest Hermes Agent (v0.17.0), Omnilimb is a Hermes plugin that lets an agent find, install, run and manage OpenClaw / ClawHub community skills — and gives it an isolated sandbox, a real Playwright browser, and multi-language runtimes. Every capability is exposed as a small, deterministic structured-JSON tool the agent calls directly, so there are zero extra LLM tokens on the execution path and no second "agent loop" burning your budget.

New in 1.0: Learn (学习技能) — point Omnilimb at any source and it distills a native Hermes skill for you. Drive it from the dashboard's visual /learn form, or just tell the agent "学习 <something>" in plain language.

The same task, without a second model call — Omnilimb executes deterministically: same result, a very different token bill.

ℹ️ Compatible with OpenClaw & ClawHub; not affiliated with them. "Omnilimb" is an independent product (omnilimb.com).

pip install omnilimb        # then: hermes plugins enable omnilimb

Why Omnilimb

• 💸 Zero extra tokens. The execution path never calls a model — the agent decides once, Omnilimb does the work deterministically and hands back JSON. Cheaper, faster, reproducible.
• 🛡️ Deterministic & safe. Retry, rollback and an optional audit log; the Docker sandbox runs with the network off by default. Path-traversal and zip-slip guarded.
• 🧠 Convert & learn. Turn any skill into a native Hermes skill — or learn a new one from a single sentence.

No lock-in and no phone-home (search talks only to the market you pick), runs with or without Node, and works across multiple markets — see Pick your market and Pick your backend below.

What it does

A small set of structured-JSON tools your agent calls directly — grouped into three jobs:

🧩 Run community skills — find & install by slug / git repo / local path (verified), run a skill's entrypoint deterministically (retry + rollback, no second AI loop), discover across markets with a per-skill health check (体检), and manage everything installed (view/edit, smoke-test, import/export). Works with or without Node via two backends.

🛡️ Execute safely — an isolated Docker sandbox (network off by default, instant rollback), Playwright browser automation via a structured action list, multi-language runtimes (python / node / bash / ruby / go), and an optional JSONL audit log.

✨ Convert & learn · new in 1.0 — convert an installed market skill into a native Hermes skill (deterministic or AI-curated), or Learn one from any source (a path, URL, or pasted text). A deterministic, no-LLM Skill butler (技能管家) health-checks, recommends and diagnoses your skills, with Learn built in.

Every tool — free in 1.0

In 1.0 every tool is free — no tiers, no license. Omnilimb registers these structured-JSON tools the agent can call:

Convert & learn — turn anything into a native Hermes skill

Two free tools take you from "found a skill" to "Hermes has it natively":

• claw_skill_to_hermes converts an already-installed market skill into a native Hermes skill under <HERMES_HOME>/skills/<name>/. Pick deterministic (pure, offline, reproducible) or ai_curated (a configured OpenAI-compatible model rewrites the docs, with a deterministic fallback).
• claw_skill_learn widens the intake to any source — a local path, a URL, or pasted text/notes — and authors a native skill from it. It's the open-ended equivalent of /learn.

Both run a structural validation loop, write transactionally, and are idempotent (re-running an unchanged source is a no-op, matched by source hash). Output lands in <HERMES_HOME>/skills/<name>/ and loads like any native skill. Drive Learn from the dashboard's visual 技能管家 → 学习技能 (/learn) form, or just tell the agent in plain language — "learn <source>" / "学习 <来源>". Everything you convert or learn shows up together in the dashboard's My skills (我的技能) library, tagged by origin (转换 / 学习).

Quickstart

As a pip package:

pip install omnilimb               # core
pip install "omnilimb[browser]"    # + Playwright
playwright install chromium        # one-time browser download
hermes plugins enable omnilimb

As a directory plugin (simplest):

cp -r omnilimb ~/.hermes/plugins/omnilimb
hermes plugins enable omnilimb

Verify inside a session:

/exo doctor

Try it locally — no Hermes, no GUI

The plugin is a headless engine; the way to feel it is to call its tools and read the JSON they return:

python scripts/demo.py doctor                  # backend status
python scripts/demo.py search github 5         # live ClawHub search
python scripts/demo.py runtime python "print(6*7)"
python scripts/demo.py sandbox "echo hi"
python scripts/demo.py menu                    # interactive

Pick your market

Switch the skills marketplace with omnilimb.market (or OMNILIMB_MARKET):

Add more under omnilimb.markets in ~/.hermes/config.yaml (each is {id, type, base_url, label} where type is one of clawhub | skillhub | clawhub_mirror | skillsmp). A new adapter class in omnilimb/registries.py is all it takes to support a new kind of market.

Pick your backend

Set omnilimb.backend in ~/.hermes/config.yaml (or OMNILIMB_BACKEND):

Dashboard UI (optional)

A dependency-free web UI ships in dashboard/ for the Hermes dashboard (rebuilt for the v0.17.0 plugin SDK; bilingual EN / 简体中文). After enabling the plugin and restarting hermes dashboard, an Omnilimb tab appears (after Skills), with these sub-tabs:

• 技能管家 (Skill butler) — the default view: a deterministic, no-LLM butler that health-checks, recommends and diagnoses your skills (and scans the audit log) from quick actions or a typed command — with the 学习技能 (learn-from-any-source) form built right in.
• Search — discovery across markets, plus a Discover mode (recommended / rising / hot / newest leaderboards + category filter), a per-skill health check (体检), and a one-click 全部体检 (health-check all) that scores every result (0–100 + grade) before you install.
• Installed — each skill expands to a detail view: 体检 health score, readiness (binaries + API keys), credentials, SKILL.md view/edit, smoke test, a one-click Update (with live progress), → Hermes convert, and export/import.
• My skills (我的技能) — the native skills you've converted and learned, gathered in one library and tagged by origin (转换 / 学习), with a front-matter view and guarded uninstall.
• Favorites — bookmarked skills.
• 审查 (Audit) — the optional JSONL audit log.
• Settings — backend / market / sandbox / cache / paths, with a compact overview.

The UI follows the active dashboard theme and language automatically.

See it in action

Real screenshots from the Omnilimb dashboard tab — the deterministic skill butler, market search with one-click 全部体检 (health-check all), installed-skill management, and the 学习技能 (Learn) form (click any image to enlarge):

技能管家 (Skill butler) — a deterministic, no-LLM butler: health-check, recommend, diagnose, plus Learn.	Search — one-click “Health-check all” scores every result (0–100 + grade) before you install.
Installed — health, credentials, smoke-test, Update & → Hermes convert.	学习技能 (Learn) — distill a native Hermes skill from a path, URL or text.

🌐 Project site: omnilimb.com

Configure (~/.hermes/config.yaml)

omnilimb:
  backend: auto            # auto | cli | native
  market: clawhub          # clawhub | skillhub | clawhub-cn | skillsmp
  sandbox_enabled: true
  sandbox_image: "python:3.12-slim"
  sandbox_network: false
  default_timeout_s: 120
  max_retries: 2
  rollback: true
  registry_base_url: "https://clawhub.ai"
  browser_headless: true
  audit_log: false         # write a JSONL audit log of tool calls
  cache_enabled: true      # local SQLite cache for discovery + search fallback
  discover_ttl_s: 21600    # discovery leaderboard cache TTL (6h)
  cache_max_age_s: 604800  # max staleness for offline search fallback (7d)

Settings changed from the dashboard's Settings tab are written to a separate overrides file (omnilimb.overrides.json), never to your hand-authored config.yaml. Resolution order is env > overrides > config.yaml.

Security

Third-party skills are untrusted code. Prefer claw_sandbox_exec with network: false for anything you don't fully trust. Without Docker, sandbox calls run locally and are flagged "sandboxed": false. Skill file operations and uninstall are path-traversal guarded; archive extraction is zip-slip protected. See SECURITY.md to report a vulnerability.

Development

pip install -e ".[dev,browser]"
pytest -q

See CONTRIBUTING.md for the architecture rules (the plugin never imports or modifies Hermes core, every handler returns JSON and never raises) and how to add a market or backend.

License

Open 1.0 — every tool and dashboard feature is free, under MIT. No paid tier, nothing to unlock — free, forever.

This repository is the full edition: skill discovery / install / run / sandbox / browser / runtimes, plus skill → native Hermes conversion (deterministic + AI-curated), open-ended claw_skill_learn, curated packs, auto-update, and the full dashboard (including the 技能管家 console).

MIT — see LICENSE. Not affiliated with OpenClaw / ClawHub.