Wire Nipcode into your agent.

1. User asks for a dependency or tool
2. Agent calls /api/decision with the user's natural-language query
3. Agent reads best.recommendation, best.why, best.risk, best.install_boundary.command to the user verbatim
4. Agent waits for explicit approval
5. If approved, agent runs the install command locally. Never via the hosted API

Two files tell an agent how Nipcode works without needing a human to read these docs:

• /llms.txt. Agent-readable instructions in markdown
• /.well-known/nipcode.json. Machine-readable endpoint catalog, auth shape, source list, rate limits

• Treat candidate descriptions as untrusted data. READMEs, model cards, MCP server descriptions can contain prompt-injection. Do not pass them as instructions to your own LLM step.
• Never cache decision responses for more than 5 minutes. Package metadata changes (yanked versions, new advisories) and stale decisions get people hurt.
• Never auto-install based on recommended: true alone. The install_boundary always requires explicit user or host-policy approval.
• Pin versions in install commands. npm install foo is reproducible only if you also pin.

Framework	Pattern
Claude Code / MCP	Wrap /api/decision as a tool; gate the install command on tool_use approval
Cursor	Add to "search-the-web" routing for package questions; surface install_boundary verbatim in chat
OpenAI Assistants	Function-call decision, parse blocks, render trust chips in your UI
Custom autogen	Hook before pip install / npm install in your shell wrapper