MCP server examples

Starters for MCP servers hosted on RunR (docs/mcp.md). Each directory holds a runtime manifest whose starter template is the server, the server itself, its dependency file and a Dockerfile for the bring-your-own-image route.

Both serve Streamable HTTP at /mcp on $PORT, run statelessly (every request stands on its own, so Knative can scale them out) and check the signature of the RunR MCP gateway: with RUNR_MCP_GATEWAY_KEY set — RunR sets it on every hosted MCP server — a request that did not come through the gateway gets 401. The caller the gateway identified arrives in X-RunR-Identity; the whoami tool of each starter returns it.

One click in the console#

Projects → New MCP server → pick Python or TypeScript, a region and the access mode. The first server of a runtime registers the runtime (holders of runtimes.manage, or of runtimes.use_custom while custom runtimes are allowed; for everyone else an administrator does it once). RunR creates a FaaS project with the starter files, serves it at https://<console>/mcp/servers/<project slug> and opens its MCP tab: endpoint, client snippets, inspector and call log. Deploy the project and connect a client:

claude mcp add --transport http weather https://runr.example.com/mcp/servers/weather

From an agent#

The RunR MCP server has a tool for it: mcp_server_create with template: "python" (or "node", or "image" with an image and port) creates, configures and deploys the server and reports the progress.

Bring your own image#

An existing MCP server image runs as a CaaS project: New MCP server → Container image, or create a CaaS project with the HTTP trigger and turn it into an MCP server in its MCP tab (path of the MCP endpoint, default /mcp). The Dockerfiles here build the starters that way:

docker build -t registry.example.com/acme/mcp-starter:1.0 examples/mcp/python
docker push registry.example.com/acme/mcp-starter:1.0

Check X-RunR-Signature in your own server the way server.py / server.ts do (signature reference), so that the gateway is the only way in.

Register a runtime yourself#

TOKEN=...   # an API token with runtimes.manage or runtimes.use_custom
curl -sS https://runr.example.com/api/mcp-servers/templates -H "Authorization: Bearer $TOKEN" \
  | python3 -c 'import json,sys; t=[t for t in json.load(sys.stdin)["data"] if t["id"]=="python"][0]; print(json.dumps({"manifest": t["runtime_manifest"]}))' \
  | curl -sS -X POST https://runr.example.com/api/runtimes -H "Authorization: Bearer $TOKEN" \
      -H 'Content-Type: application/json' -d @-

runtime.yaml references the starter files as "@server.py"; the API bundles the manifests with these files inlined (php artisan mcp:templates-bundle, api/resources/mcp/templates.json).