MCP Server

apps/mcp-server exposes the same tool catalog the internal AI copilot uses (look up/reply to/change status of/assign/apply a macro to/escalate a ticket) to your own MCP-compatible agent — Claude Desktop, an n8n flow, your own script. It's how you connect an external AI agent to Seredina without reimplementing any business logic or bypassing any authorization controls.

Two transports

Chosen with MCP_TRANSPORT (defaults to stdio):

stdio

A separate process from api/worker — not part of a plain docker compose up. The client spawns it per session and owns its stdin/stdout directly, the same way Claude Desktop spawns any other local MCP server. The tenant is resolved once at startup, from SEREDINA_API_KEY.

# 1. Create an API Key from Administration -> API Keys in the console.

# 2. Build the image:
docker build -f infra/docker/Dockerfile.mcp-server -t seredina-mcp-server .

# 3. Point your MCP client at:
docker run -i --rm \
  -e SEREDINA_API_KEY=<your key> \
  -e DATABASE_URL=... \
  -e ENCRYPTION_KEY=... \
  seredina-mcp-server

In development, running apps/mcp-server directly with tsx/node (see the dev script in apps/mcp-server/package.json) skips the image build step.

http

A genuinely always-on network service (Streamable HTTP, deliberately stateless), for a remote or long-running agent instead of a locally spawned one. Every request carries its own Authorization header with a Bearer API Key — the tenant is resolved per request, never cached server-side.

docker compose --profile mcp up mcp-server-http

Deliberately profile-gated — a plain docker compose up never starts it. You can also run it directly with MCP_TRANSPORT=http in apps/mcp-server.

Authorization

Every tool call that changes data goes through the tenant's Autonomy Policy (Settings → AI Agent Activity in the console): tools not on the auto-execute allow-list wait there for a human to approve or reject — and every call, whether auto-executed or not, lands in the same audit log. Your external agent has exactly the same limits as the internal copilot, not a separate path with fewer controls.

Tool catalog

The same set the internal copilot uses for its autonomous mode — look up a ticket, add a reply, change status, assign, apply a macro, escalate. See AI Copilot in the user guide for how the human-approval side works from the console.