The handoff

Connect an implementer

There are two ways to hand your approved model to an AI. Paste it — copy the whole model as a single prompt, no setup at all. Or connect it — Amy serves the project over MCP, and any MCP-capable agent pulls the .aim model plus the change records (the forward graph-diff of every accepted change, AIM v5.2 §16.4) live, so when the model evolves the implementer pulls what changed and propagates it.

The quick way: one prompt, no setup

In Amy, open Share → Export and copy the model as an agent prompt. It is self-contained — the instructions, a condensed language reference, and every .aim file inline — so it works in Claude, a desktop agent, or anything that reads text. Two flavors:

Build The agent writes the software that realizes the model, plus the tests that hold it to the model's requirements.
Operate The agent runs the process instead: it keeps the records as files, follows each Flow's steps, watches the deadlines, and routes every human step to the person who owns it.

Either way the agent starts by summarizing the model back to you, so you can check its understanding before it acts — and it never edits the model. Gaps come back as questions for you to decide; the model stays yours.

What counts as an implementer?

Anything that turns the model into results. The connection is identical — only the output differs.

Coding agent Claude Code, Cursor, Gemini CLI… pulls the model and writes the app plus its tests.
Operations agent OpenClaw or any agent runtime — performs the mapped process: triggers fire, approvals block, human steps go to their people.
Writing agent Drafts the script or manuscript from the story model — and keeps its setups and payoffs honest.

Connect over MCP when the model keeps changing and the implementer should always work from the current one. Paste the prompt for a one-shot handoff or an agent without MCP.

What the implementer can pull

1Get a connect token

In Amy, open the project and choose Share → Implementer access → New token. You'll get a ready-to-paste command with your base URL and token already filled in. The token is shown once — copy it then. It grants read-only access to that one project; revoke it any time from the same panel.

2Add it to your implementer

Claude Code

Run the command from the Share dialog (this is its shape):

claude mcp add --transport http amy "https://amybuilds.com/mcp" --header "Authorization: Bearer <TOKEN>"

Add -s user to make it available across all your folders. Avoid -s project — that writes the token into a committed .mcp.json. Then run /mcp in a session to confirm amy is connected.

OpenClaw, Cursor, Windsurf, or any mcp.json

{
  "mcpServers": {
    "amy": {
      "url": "https://amybuilds.com/mcp",
      "headers": { "Authorization": "Bearer <TOKEN>" }
    }
  }
}

For clients that don't support custom headers, put the token in the path instead: https://amybuilds.com/mcp/<TOKEN> with no headers.

Replace amybuilds.com with your own host if self-hosting (e.g. http://localhost:8787 for local development). The Share dialog always fills in the correct base URL for you.

3Verify & use

In a Claude Code session run /mcp (it should list amy), or just ask your implementer:

“List your MCP tools, then call get_bootstrap, then get_intent.”

The recommended loop for any implementer:

  1. get_bootstrap — reads the spec URL and the workflow; the implementer self-configures.
  2. get_intent — pull the .aim model and produce the result: code + tests, a running process, a drafted script.
  3. get_change_records(since: "<last semver>") — on every rebuild, pull only what changed and propagate it as targeted edits.

Smoke-test the endpoint (optional)

curl -s -X POST "https://amybuilds.com/mcp" \
  -H "Authorization: Bearer <TOKEN>" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}'