For agents

keelyard is meant to be driven by agents. Everything a person can do from the CLI, an agent can do through MCP, and both return JSON.

What init installs

agent	hooks	MCP	instructions
Claude Code	.claude/settings.json	.mcp.json	CLAUDE.md section, .claude/skills/keelyard/SKILL.md, /keelyard:status, /keelyard:prove, /keelyard:undo
Codex	.codex/hooks.json	snippet for ~/.codex/config.toml	AGENTS.md section
OpenCode	.opencode/plugin/keelyard.js	opencode.json	AGENTS.md section

The instruction sections sit between <!-- keelyard:start --> and <!-- keelyard:end -->. init replaces only that block, and uninstall removes only that block. The rest of the file stays as you wrote it.

The loop an agent follows

1. proof_requirements(items=[...])
2. project_map, find_symbol, outline, then read_file
3. edit (snapshots happen automatically)
4. run_compact for tests and builds
5. proof_run_tests(phase="after"), then proof_check for each item
6. proof_status → {"proven": true}

All 20 tools are listed, with their schemas, in the MCP reference.

Self-test

keelyard selftest plays the agent. It sends the same hook JSON Claude Code sends and does everything else through the MCP server over stdio. In a temporary project it records requirements, reads code, searches memory, runs a swarm task against a local stub model, makes a bad edit, gets blocked by the gate, rewinds, fixes the bug, proves it and is let through. It needs no network and no API key.