Platform vs custom
| Type | Who publishes | How you install |
|---|---|---|
| Platform catalog | AgentRuntime operators | MCP → Add connectors → Add instance on a catalog row |
| Custom (third-party) | Your workspace | MCP → Add connectors → Add connector → register URL you operate |
mcp_servers row scoped to your workspace. You still wire connections on Instance config profiles when tools need credentials.
Requirements
Your MCP server must:- Speak MCP over HTTP at a stable URL (for example
https://tools.example.com/mcp) - Expose tools with JSON input and output schemas
- Be reachable from AgentRuntime’s runtime network (public HTTPS or allowlisted private endpoint)
- Return at least one tool on discover — empty catalogs fail registration
agentruntime-mcp for Python, agentruntime-mcp-go for Go) provide router helpers compatible with platform discovery.
Register in the Console
Add connector
In the marketplace modal, click Add connector (not Add instance on a platform row). This opens the registration wizard.
Discover tools
AgentRuntime calls your server and imports the tool catalog and config schema. If discovery returns zero tools, check that the URL is correct and the MCP endpoint is running.
Review catalog
Review discovered tools (name, input/output shape). Use Hold to inspect the list, or wait for the countdown to continue automatically.
Publish and activate
The wizard publishes version
1.0.0, runs validation checks, and creates a workspace instance.Configure credentials
Open Instance config (
/mcp/instances/{instance_id}). On the active profile, wire a connection if tools need API keys or OAuth tokens./mcp/{server_id}/validate.
Register via API
Requires project_contributor:| Step | Endpoint |
|---|---|
| Discover tools | POST /v1/mcp/servers/{id}/discover |
| Publish version | POST /v1/mcp/servers/{id}/versions with tools, config_schema, activate: true |
| Validate | POST /v1/mcp/servers/{id}/validate |
| Add instance | POST /v1/mcp/instances with { "server_id": "…" } |
Auth and connections
The registration wizard defaults toauth_mode: "token". Credential delivery depends on your server:
- Connection-backed — Store secrets in Connections, bind on the instance’s active config profile (
source_spec). Preferred for API keys and OAuth refresh tokens. - Server-side auth — If your MCP server validates its own tokens at the HTTP layer, you may not need per-instance wiring (still validate in a test run).
Troubleshooting
| Symptom | Fix |
|---|---|
| No tools discovered | Confirm URL, TLS, and that the MCP handshake returns tools. Test with curl from your network. |
| Validation failed | Open /mcp/{server_id}/validate for per-check messages (reachability, schema, smoke call). |
| Tools missing in Workflow Studio | Ensure the instance exists and the active config profile is set. Save and re-open the workflow. |
| 401 / 403 at runtime | Wire the correct connection on Instance config or fix server-side auth. |