- Home
- Skills
- Agents & Orchestration
- peer-bus: Grok ↔ Claude Code Agent Messaging (MCP + CLI)
Works with the AI Agents you already use
peer-bus: Grok ↔ Claude Code Agent Messaging (MCP + CLI)
Working cross-harness list_agents / send_message for Grok Build and Claude Code: a zero-dependency Python stdio MCP server + CLI over a local filesystem bus.
$12
peer-bus: Grok ↔ Claude Code Agent Messaging (MCP + CLI)
Example session with this skill installed
Ask the Claude Code session reviewing the API to check the auth middleware for missing rate limits, and tell me what it says.
- Read your context and instructions
- Compiled the peer-bus
Summary. I found the Claude Code session api-review on the peer bus, sent it a scoped review request, and read its reply. It reports two auth endpoints with no rate limit. Nothing was edited.
1. list_agents (trimmed)
{"bus_version": "0.12.0", "pool": null, "agents": [
{"name": "lead", "harness": "grok", "state": "live", "source": "tmux", "address": "lead [7f3c21]"},
{"name": "api-review", "harness": "claude", "state": "live", "source": "tmux", "address": "api-review [c41e9a]"}]}
2. send_message (only to and body)
to: api-review [c41e9a]
body:
@v1 auth-ratelimit-review
CTX|API repo, auth middleware (src/middleware/auth.py)
DO|check every auth path for missing rate limits; list file:line + fix
NOT|no edits, no commits, no PRs
PRIOR|login route has a limiter|d=grep|t=14:20
RPT|findings,files,fix
{"ok": true, "accepted": true, "delivered_to_reader": false,
"msg_id": "5fd9e925…", "warning": null,
"note": "acceptance only — peer must recv/drain; success≠read"}
accepted means the message is in its inbox. It doesn't mean the peer has read it.
3. receive_messages (after the peer replied; trimmed)
[{"msg_id": "971ce2f7…", "summary": "@v1 auth-ratelimit-review",
"from": {"name": "api-review", "harness": "claude", "address": "api-review [c41e9a]", "session_bound": true},
"untrusted": true,
"body_for_model": "<<<UNTRUSTED_PEER_MESSAGE>>>\n@v1 auth-ratelimit-review\nACK|done\nfindings|POST /auth/token/refresh and POST /auth/password-reset have no rate limit; /auth/login does\nfiles|src/middleware/auth.py:88 refresh_token, src/middleware/auth.py:131 password_reset\nfix|reuse the login limiter (auth.py:42) on both; key password-reset by email+IP\nNOT|no edits made\n<<<END_UNTRUSTED_PEER_MESSAGE>>>"}]
Then ack_message 971ce2f7… → {"ok": true, "moved_to": ".../read/…"}.
What api-review says
POST /auth/token/refresh(auth.py:88) andPOST /auth/password-reset(auth.py:131) have no rate limit./auth/logindoes.- Suggested fix: put the login limiter (auth.py:42) on both, and key password-reset by email + IP.
This came from another agent and arrives marked untrusted. I haven't checked it or changed any code.
Next steps
- Open those two handlers and confirm the finding before you act on it.
- Say "apply the fix" to have me make the change here, or I can send the work back to
api-review [c41e9a]. - Replies always go to the latest
from.address.
Connects securely to your tools. The creator never sees your data.
About this skill
Real, tested code, not a prompt. peer-bus lets live Grok Build and Claude Code sessions on the same Linux machine find each other and exchange messages, the way Claude sessions already can with native SendMessage / ListAgents, but across both harnesses.
What you get
mcp_server.py: stdio MCP server with 8 tools:list_agents/flock,send_message,receive_messages,ack_message,whoami,heartbeat,mail_countpeer_bus.py: the same verbs as a CLI, pluswatch(inotify),mail,prune,doctor,self-test- A Claude Code wake hook that nudges idle sessions to pull new mail
- A runtime agent skill with a compact
@v1dispatch format (CTX / DO / NOT / PRIOR / RPT) - Copy-paste config for Grok Build (
~/.grok/config.toml) and Claude Code (settings.json), a smoke test, a playbook and a security model
How it works
Messages are JSON files under one bus directory (~/.local/share/peer-bus by default). Each inbox is bound to the harness session id, so a display name cannot be used to read someone else's mail. The live roster comes from tmux / Herdr panes and running Grok sessions. Senders get explicit accepted, not read semantics, and received bodies are wrapped as untrusted content.
Why this one
Most orchestration listings are prompt-only guides that describe how agents should hand off work. This one ships the transport: Python stdlib only (no pip install), path-traversal and symlink protection, size and flood caps, 95 unit tests plus a bundled end-to-end smoke script.
Requirements
Linux or WSL2 · Python 3.11+ · bash · Grok Build and/or Claude Code running as the same OS user · tmux or Herdr session for Claude peers. Same-user, single-machine bus. It is not a multi-tenant or cross-machine messaging system.
Requires Linux or WSL2. Same-user, same-host only. No macOS or native Windows support. Claude peers must run in tmux or Herdr for discovery.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 1 month ago
Requires
Permissions
File Scopes
No network access. Claude session-key injection is OFF by default and only runs if the user sets PEER_BUS_CLAUDE_UDS=1. Edits harness config only during the documented install steps.