Hermes Remote MCP Setup
When to Use
Use this skill when adding, authorizing, troubleshooting, or verifying a remote OAuth MCP server in Hermes.
Procedure
- Preserve the requested server identity and endpoint. Use the exact URL supplied by the user first. If it returns a non-MCP response, consult the vendor's official documentation or endpoint metadata before trying an alternate endpoint; never silently substitute a guessed URL.
- Add OAuth-capable servers with the native CLI. Use:
Addhermes mcp add <name> --url <endpoint> --auth oauth--connect-timeout 120when authorization or tool discovery is slow. - Run interactive setup through a PTY. The initial OAuth flow may open a browser and pause for authorization. Keep the process alive, tell the user to complete authorization in the browser, and never request or accept OAuth codes in chat.
- Enable tools explicitly. When Hermes lists discovered tools, select all only when the user asked to install the server generally; otherwise select the requested subset. In PTY flows, a raw
y\rmay be needed after a normal submit if the selection prompt remains open. - Verify before reporting success. Run:
Report the endpoint, auth method, connection status, and exact discovered/enabled tool count.hermes mcp list hermes mcp test <name> hermes config get mcp_servers.<name> - State activation scope. Native MCP discovery is startup/session scoped. Tell the user to run
/reload-mcpor start a new session before using newly added tools in an already-open conversation.
Pitfalls
- Do not save a server after a failed endpoint probe unless the user explicitly wants a deferred, unverified configuration; a 403 HTML page or an SSE content-type mismatch is evidence that the URL is not the usable MCP endpoint.
- Do not claim OAuth setup succeeded merely because Hermes wrote configuration; require a successful authenticated
hermes mcp test. - Keep credentials, authorization codes, bearer tokens, and full OAuth URLs out of memory, skills, and chat; they are secrets or short-lived session data.
- Treat vendor endpoint redirects as a discovery problem, not a reason to normalize the user's URL silently; preserve the original URL in the report and name the verified alternate endpoint.