| name | claw-hermes |
| description | Bridges Hermes Agent (closed learning loop, FTS5, Honcho) and OpenClaw (~24 channels, Voice Wake, Live Canvas) into one personal AI OS. Use when the user wants to (1) review PRs / triage issues / classify AI-slop on GitHub repos, (2) push events to messaging channels (Telegram, WhatsApp, iMessage, Discord, Slack, Matrix, WeChat, …), (3) bridge memory between channels (iMessage remembers Slack), or (4) ask "is there a way to use Hermes and OpenClaw together?" — claw-hermes is that bridge. |
| version | 0.2.0 |
| license | MIT |
| author | Albert Kamalov |
| agentskills_version | 1.0 |
| runtimes | {"hermes":{"entrypoint":"python -m claw_hermes.cli","capabilities":["github","memory.recall","memory.write","subprocess"]},"openclaw":{"entrypoint":"claw-hermes","capabilities":["github","channels.send","network"]},"both":{"requires_capabilities":["github","memory.recall","channels.send"]}} |
| keywords | ["github","agent","hermes","openclaw","automation","messaging","ai"] |
| homepage | https://github.com/bertbertov/claw-hermes |
claw-hermes — personal AI OS bridge
claw-hermes is the connective tissue between two of the largest open AI assistant projects:
- Hermes Agent (Python, Nous Research, 119k stars) — closed learning loop, skill autogeneration, FTS5 session memory, Honcho dialectic user modelling, RL trajectory generation
- OpenClaw (TypeScript, 365k stars, 174k Discord) — ~24 channel surfaces incl. iMessage, WhatsApp, WeChat, Matrix, Voice Wake, Live Canvas
Tagline: "Letta with hands and ears."
The thesis: a self-hosted agent that learns from you AND lives in every messenger you read. No other product (open or closed) currently combines a closed learning loop + 24-channel native delivery + self-hostable.
When to invoke this skill
- "review PR 123 in "
- "classify this PR as slop or legit"
- "give me a digest of open PRs in "
- "what channels would a CI failure route to?"
- "send the latest release notes to my team across iMessage and Slack"
- "can I use Hermes and OpenClaw together?"
- Any GitHub event → messaging channel orchestration question
- Any question about marrying Hermes + OpenClaw
Prerequisites
pip install claw-hermes
gh auth status
Optional but recommended for full functionality:
pip install hermes-agent — enables real PR reviews (otherwise a deterministic skeleton is used)npm i -g openclaw && openclaw onboard --install-daemon — enables real channel delivery
Core commands (v0.1)
| Command | What |
|---|
claw-hermes init | Write default routing config to ~/.claw-hermes/config.yaml |
claw-hermes status | Show wiring: gh auth, hermes availability, OpenClaw reachability |
claw-hermes route <event> | Show which channels a GitHub event would route to |
claw-hermes pr-fetch <repo> <pr#> | Fetch a PR via gh (read-only) |
claw-hermes pr-review <repo> <pr#> | Generate a review digest (uses Hermes if installed) |
claw-hermes pr-review <repo> <pr#> --deliver --dry-run | Show channel routing without firing |
claw-hermes hermes-probe | Read-only Hermes availability check |
claw-hermes openclaw-probe | Read-only OpenClaw gateway HTTP probe |
v0.2 commands (shipped)
| Command | What |
|---|
claw-hermes slop-classify <repo> <pr#> | Classify a PR as human / ai-slop / ai-assisted-legit (Hermes if installed; deterministic heuristic fallback) |
claw-hermes slop-classify <repo> <pr#> --record | Persist the verdict to the cross-repo signature store at ~/.claw-hermes/signatures.db |
claw-hermes slop-classify <repo> <pr#> --json | Machine-readable output |
claw-hermes slop-classify <repo> <pr#> --deliver --dry-run | Route verdict via OpenClaw pr_review_requested rule |
claw-hermes skill lint <path> | Validate a SKILL.md manifest (file or directory) against agentskills.io v1.0 + the runtimes: extension |
claw-hermes skill new <name> | Scaffold a new dual-runtime skill at <name>/SKILL.md and lint it automatically |
claw-hermes skill list <dir> | Discover all */SKILL.md skills under a directory and print a one-line summary each |
Slop-classifier signals (v0.2)
The heuristic fallback fires named signals that get persisted with each --record-ed verdict.
Use these names if you want to inspect or filter the SQLite store directly:
ai_phrase_present — body or title contains "as an ai", "generated with", "co-authored-by: claude", etc.emoji_heavy_body — body is at least 40 chars and >5% emoji.suspiciously_round_size — additions+deletions is a round multiple of 100 and >=100.mass_rename_diff — additions == deletions, both >0, across many files.hallucinated_import — diff imports a non-existent stdlib-shaped module (e.g. json_helper, requests_async).no_tests_on_large_pr — diff touches >100 lines but no tests/, _test., or /spec paths.templated_phrasing — body contains 4+ section markers like ## Summary, - [x].
Coming in v0.3
| Command | What |
|---|
claw-hermes triage <repo> | Per-contributor Honcho model surfaces "this contributor's last 5 PRs all merged" |
claw-hermes verify | Walk every pipeline stage; refuse to claim success without terminal-channel ack |
Workflow patterns
Pattern: review-then-route
claw-hermes pr-review <repo> <pr#> to generate a digest- Pipe to
--deliver --dry-run first so the user sees where it would go
- Drop
--dry-run once they confirm
Pattern: status-before-anything
Always run claw-hermes status before claiming a delivery worked — it reports whether OpenClaw is actually reachable. Don't trust silence.
Pattern: graceful degradation
If Hermes isn't installed, the review falls back to a deterministic skeleton with PR metadata. Tell the user this happened so they can install Hermes if they want the AI review.
Verification before claiming done
claw-hermes itself codifies a verify-end-to-end rule. After any delivery, check the delivered_to and failed fields in the output. Don't say "sent" just because the command exited 0 — dry_run=True always exits 0 even though no message left the box.
Skill provenance + roadmap
This skill ships with the claw-hermes repo at skill/SKILL.md. To install:
mkdir -p ~/.claude/skills/claw-hermes
cp -r /path/to/claw-hermes/skill/* ~/.claude/skills/claw-hermes/
For the v0.2 → v1.0 trajectory (federation, MCP, cross-runtime self-improvement), see ROADMAP.md. For the federation design, see ARCHITECTURE.md.
