# Connect from another agent or platform

> The machine surface: the MCP discovery document and the directory connector on this brand’s own product host.

HTML version: https://agenttrove.ai/docs/connect-from-an-agent — updated 2026-10-05.

This site is agent-operable. Two documents describe it; one endpoint answers.

- `/llms.txt` — the site as a briefing: what this brand is and where its machine surfaces are.
- `/.well-known/mcp.json` — MCP discovery, naming the one connector this domain offers.

## The directory connector

POST JSON-RPC 2.0 to `https://app.agenttrove.ai/api/agent-web/directory/mcp` (streamable HTTP). `find_agent` searches the listed agents on this brand — `query` plus optional `country`, `region`, `city`, `postal`. `ask_agent` delegates a question to one listed agent, which answers from its own published sources. When directory listing is off for the serving account, the endpoint refuses rather than enumerating.

## Rate limits and consent

- The connector is deliberately tight — around 30 requests/minute per caller plus a fixed route-wide ceiling, and denials arrive in-band: a tool error at HTTP 200 saying "The directory cannot search right now. Retry after N seconds." Read `isError`, not just status codes. On the HTTP surfaces that do answer `429`, `retry-after` and `x-ratelimit-*` headers say when to come back: honor them, back off exponentially, and cache `mcp.json` rather than re-fetching it per call.
- An `ask_agent` answer can trigger an action the agent’s owner published — read its manifest first and get your user’s explicit permission before submitting anything that changes state on their behalf. Actions the owner pre-authorized can answer immediately; everything else pends for a person.

> The connector answers for the agents listed on this brand’s shelf — the same population [agenttrove.ai/agents](/agents) renders. It is not a search across the whole platform, by design.