The AI Agent Brief — GIVE-THIS-TO-YOUR-AI-AGENT.md

Glassy's best setup document for AI agents is GIVE-THIS-TO-YOUR-AI-AGENT.md in the installer repo. If you want an agent (Claude, Cursor, Windsurf, OpenCode, Hermes, or any MCP client) to operate your appliance without a human explaining it, start there — not here.

What it is. A self-contained onboarding brief, versioned to the running build — its header states the release it was verified against. It covers: connecting to the MCP surface and reading /api/capabilities before writing, the full tool contract (including which fields a write can ignore and what warnings[] / degraded[] mean), durable agent memory and its lifecycle, the owner review loop, Obsidian vault lanes (bridge, proxy, and the agent-as-vault-provider path), and the ops facts an agent needs on a live appliance.

Why it is machine-checked, and why that matters. scripts/check-doc-tool-counts.js, scripts/check-mcp-claims.sh, and the docClaimFreshnessGuard test compare every tool / prompt / resource count and canvas claim in the brief against the actual code, so the document cannot silently fall behind the product. A from-scratch integration that follows only the brief has needed zero workarounds since v2.43.0.

How to use it.

1. Clone or bookmark the raw brief: GIVE-THIS-TO-YOUR-AI-AGENT.md (raw).
2. Give it to your agent as the first document of the session.
3. The agent's first verified action should be GET /api/capabilities — the brief teaches the read-before-write posture every other step assumes.

The pages in this section describe the product surface; the brief describes how to operate it. When you find a disagreement between the two, the product is what runs — but please file it, because one of them being wrong is exactly the defect class those machine-checks exist to prevent.