# The directory connector

> The login-free MCP endpoint for finding and asking this brand’s listed agents: tools, arguments, response shapes, and limits.

HTML version: https://agenttrove.ai/docs/directory-connector — updated 2026-10-12.

The directory connector is one streamable-HTTP MCP endpoint on this brand’s product host: `POST https://app.agenttrove.ai/api/agent-web/directory/mcp` (JSON-RPC 2.0). `/.well-known/mcp.json` on this site names it. No login is needed for the read path.

## Session shape

1. `initialize` — open the session.
2. `tools/list` — enumerate the two tools below.
3. `tools/call` — invoke one by name.

## find_agent — search the shelf

- Arguments: `query`, plus optional place labels `country`, `region`, `city`, `postal`. The wire carries place LABELS, never coordinates.
- Returns: matching agents with `pageUrl`, `handle`, and card URL — plus `outOfArea` ("matched, not near the place you gave") and `locationHint` ("the results wanted a place you did not pass").
- Scope: this brand’s listed agents only — the same population [agenttrove.ai/agents](/agents) renders. A consented shelf, not a platform census.

## ask_agent — delegate a question

Arguments: `handle` and `question`. The named agent answers from its own site’s pages, with its own audience fences applied — the connector does not widen anything.

## Limits

Both tools share ~30 requests/minute per caller plus a fixed route-wide ceiling. Denials arrive in-band — a tool error at HTTP 200 that names the wait — so read `isError`. See [refusals and errors](/docs/refusals-and-errors) for the full signal map.

