opencoat-skill

The 5-minute install skill for OpenCOAT — the Open Concern-Oriented Agent Thinking Runtime.

Drop this skill into any agent that supports the agent-skill convention (Cursor, Claude Code, Codex, …). The entire host→agent instruction can be a single line (see SKILL.md §Canonical host prompt); all requirements for the agent live in SKILL.md and linked files. The skill tells the agent how to:

• install opencoat-runtime from PyPI — one pipx install opencoat-runtime puts the opencoat CLI on PATH, and pipx inject opencoat-runtime opencoat-runtime-host wires the host SDK into the same env so the demo's lazy imports succeed (opencoat-runtime-protocol comes along transitively),
• start a local daemon (zero-config HTTP on 127.0.0.1:7878/rpc) and register OS autostart via opencoat service install,
• import three dramatic demo concerns,
• close the full loop with one command — opencoat demo fires three events through install_hooks and uses the pickup API (apply_to / guard_tool_call) to fold the resulting advice back into a tiny in-script host. Three scenes print BEFORE / AFTER so users see concerns visibly change behavior on first read. opencoat demo --in-proc runs the same tour without a daemon (no setup), and opencoat demo --script-out demo_host.py dumps the equivalent Python for users who want a template to adapt for a real host. OpenClaw gateway: TS bridge in the OpenCOAT monorepo at integrations/openclaw-opencoat-bridge/ (clone ~/OpenCOAT, Step 5 in SKILL.md — not a separate repo). Then
• inspect the resulting Deep Concern Network.

Layout

opencoat-skill/
├── SKILL.md                       # main instructions (≤500 lines)
├── skill.json                     # manifest
├── inspection.md                  # read-only inspection commands
├── concerns.md                    # authoring patterns + recipe gallery
├── rules.md                       # safety rules the host must respect
├── bootstrap_daemon.sh            # post-pipx: runtime up + service install
├── DEPLOY.md                      # one-time CF Worker + DNS setup
├── worker/                        # Cloudflare Worker source (deploys via CI)
├── LICENSE                        # Apache-2.0
└── .github/workflows/
    ├── verify.yml                 # CI: parse skill.json + link integrity
    └── deploy-worker.yml          # CI: redeploy worker on skill / worker push

Install (per-agent)

Agent	Path
Cursor (personal)	~/.cursor/skills/opencoat/
Cursor (project)	<repo>/.cursor/skills/opencoat/
Claude Code	~/.claude/skills/opencoat/
Codex	~/.codex/skills/opencoat/
OpenClaw gateway (TUI / Telegram)	No skill folder required — agent must auto-run Step 5 (install-local.sh from resolved OPENCOAT_ROOT / ~/OpenCOAT). Optional: workspace/skills/opencoat/ for always-on instructions

Three install paths, pick whichever fits the agent:

1. Self-install from the website — point the agent at https://www.opencoat.ai/SKILL.md and tell it to "read the skill and follow the install instructions". The agent fetches all five skill files (SKILL.md, concerns.md, inspection.md, rules.md, skill.json) from https://www.opencoat.ai/<file>, drops them into its skills directory, then walks the user through the runtime install. This is the most "Moltbook-shaped" path and the one the website is optimized for. The Cloudflare Worker + DNS setup that powers these URLs is documented in DEPLOY.md — once for the maintainer, never for the user.
2. Clone the repo into the agent path above. Useful if you want to track skill changes with git pull.
3. Skill installer — if your agent has a built-in skill installer (Cursor, Codex), point it at this repo URL.

Versioning

This skill tracks the major version of opencoat-runtime (which ships the daemon + the opencoat CLI in one wheel as of 0.1.0). See skill.json compatible_with for the exact minimum versions of each upstream package.

Issues

• Skill content / phrasing / wiring → file here.
• Upstream runtime bugs → file at https://github.com/HyperdustLabs/OpenCOAT/issues.

License

Apache-2.0. See LICENSE.