When the Hermes Agent connects to the MCP service, if the service exits as soon as it starts, it prioritizes the command path, working directory, environment variables, and stderr logs. The MCP configuration looks like a minor JSON/YAML issue, but the real problem is often "commands cannot be found in Hermes' running environment".
First reproduce in the terminal
Copy the command and args in the MCP configuration to the same user and the same shell for execution. If the terminal can't run, Hermes won't magically run either. Common situations include Node/Python not in the PATH, the virtual environment is not active, the package is not installed, and the script path uses a relative path.
Four key checkpoints
- command once with an absolute path, e.g. python in
/usr/bin/nodeor venv. - CWD correct: Many MCP servers need to be started at the root of the project.
- env complete: API key, configuration path, proxy variables don't just exist in the interactive shell.
- Logs with stderr: The real reason is usually printed before the service exits.
Don't open all MCPs together
Enable only one MCP server during troubleshooting, make sure it starts stably, lists tools, and makes a minimal request before connecting to the next one. When multiple MCPs fail at the same time, the logs are mixed together, making it difficult to determine who is exiting.
If MCP can run in Cursor or Claude Code, but not in Hermes, it is probably because the boot environment is different. Explicitly writing paths, cwd, and env into the Hermes MCP configuration is more stable than relying on the implicit state of the shell.
