# MCP management & aggregation

> Phases 2.2 + 2.4 (dial sites: Phase 8 C2) · designs:
> [registry](design-phase-2.2-registry.md) ·
> [connect & tools](design-phase-2.4-connect-tools.md) ·
> [dial sites](design-phase-8-c2.md)

Register upstream MCP servers once; every agent tool consumes the aggregated
outlet instead of dialing each upstream itself.

## MCP server rows

Each row declares a transport (`sse` / `http` / `stdio`) plus any
`${cred:NAME}` placeholders in url / command / args / env / headers.
Credentials are resolved at connect time (see [Credentials](credentials.md)).

## Where the connection lives — dial sites

- **server** — the platform dials the upstream (SSE/HTTP) and serves it
  through the aggregated outlet.
- **client** — the user's machine dials it via the `hnx mcp serve` shim
  (SSE/HTTP/**stdio**; every referenced credential must be distributable).
- **auto** — client iff all referenced credentials are distributable (or
  none), else server. stdio can never be server-dialed.

## The aggregated outlet

- `GET/POST /mcp` (Streamable HTTP) and legacy SSE at `/mcp/sse` — PAT-gated,
  **routed by profile**: `/mcp?profile=<id>` exposes exactly that profile's
  MCP servers.
- Tools are namespaced `serverName__toolName` to avoid collisions.
- Unreachable upstreams are marked `error` and skipped — they never block
  startup or tool listing.

## Operator controls

- Live status (`connected / connecting / error / disconnected`) + tool counts
  drive the dashboard mesh and the management page badges.
- **Connect / Disconnect** buttons per server-dialed row (re-dial after a
  credential fix; drop a live connection without deleting the row).
- Expandable **tool inspection**: each upstream tool with description and
  `inputSchema` parameters, plus a refresh — in original, un-namespaced
  names.