This shows how to connect Amp — Sourcegraph's coding agent — to RAC, so Amp respects the decisions your team has already recorded. A stranger can reproduce the setup from this file alone.

RAC meets Amp on two surfaces Amp already supports natively, so there is no Amp-specific RAC code: a generated AGENTS.md (context Amp reads every session) and the lore MCP server (tools Amp queries on demand). They are complementary — one pushes the decisions into context, the other lets Amp pull the full text and relationships when it needs them.

pip install rac-core   # the `rac` CLI and the `lore` MCP server

You also need a repository with a RAC corpus under decisions/ (run decided quickstart in a fresh project, or point at this repository's own decisions/).

Amp reads AGENTS.md for project conventions and guidance. RAC generates one from your recorded decisions:

decided export decisions/ --agent-rules

This writes (or updates) AGENTS.md at the repository root, inside a managed block — anything you keep outside the block is preserved. Amp discovers it automatically: it reads AGENTS.md from the working directory up through the parent directories, and from subtrees as it opens files there. Re-run the export whenever decisions change (a pre-commit hook or CI step keeps it current); the agent-rules drift check (decided export decisions/ --agent-rules --check) fails if it falls out of sync.

Amp falls back to CLAUDE.md when no AGENTS.md is present — RAC generates that target too, so either way Amp sees the decisions.

Amp configures MCP servers under the amp.mcpServers key. Add a project-level .amp/settings.json at the repository root (a sample is in settings.example.json):

{
  "amp.mcpServers": {
    "asdecided": {
      "command": "decided-mcp",
      "args": ["--root", "."]
    }
  }
}

• Project config: .amp/settings.json in the repo root (commit it so the team shares it). Use "." for --root so it resolves to the repository.
• User config: ~/.config/amp/settings.json — use an absolute --root path there.

This exposes RAC's five read-only lore tools to Amp: get_summary, search_artifacts, get_artifact, get_related, and find_decisions. The server re-reads the corpus from disk on every call, returns structured JSON errors rather than exceptions, and never writes to the repository.

Use the bundled grounding demo to see the difference a connection makes. The examples/guide/ demo runs the same coding task twice — once with no MCP server and once with lore connected — and shows the unconnected agent violating a recorded decision that the connected one respects. Point Amp at it the same way (its prompt and corpus are client-agnostic), or run decided-mcp --root examples/guide and ask Amp to implement the task in examples/guide/task/.

RAC supplies context and enforces after the edit — it does not intercept Amp's edit loop (ADR-067). Whatever Amp writes is checked by decided validate and decided relationships --validate (and the GitHub Action / pre-merge gate) the same as any other contributor; the trust boundary is human PR review and CI, not the agent. The per-edit decided hook is specific to Claude Code, so with Amp you rely on the CI gate rather than a live pre-edit block — Amp's loop stays untouched and adds no latency.