Zotero Mcp Write Mode Config Change 2026 07 02

The CLI hermes config set accepts most values, but the literal string ${ZOTERO_API_KEY} (the env-var-reference syntax that keeper's MCP config uses) was rejected by the CLI parser on 2026-07-02 (silent traceback, no value written). The workaround: sed -i '/ ZOTERO_LIBRARY_ID: <id>/a\\ ZOTERO_API_KEY: ${ZOTERO_API_KEY}' <profile config> directly edits the yaml while preserving indentation and quoting. Cleaner than the patch tool, which refuses to write to ~/.hermes/**/config.yaml (security-sensitive).

zotero-mcp is stdio-spawned by the gateway. When the gateway loads MCP servers, it reads ~/.hermes/profiles/build/config.yaml once at spawn time, then starts zotero-mcp serve --transport stdio with the resolved env vars. Within a single TUI session, the LLM is bound to the MCP child that was spawned at session start. Mid-session config edits (like the hermes config set + sed sequence done on 2026-07-02) require the gateway to respawn the MCP child for the new config to take effect.

The gateway uses --replace mode and is launched by launchd via KeepAlive. kickstart -k sends SIGKILL, which launchd treats as "service crashed, restart it." launchd then re-runs the program, which includes the Python gateway, but the gateway may not re-spawn MCP children until a new turn. The MCP children seen in ps -ef (e.g., PID 19287's zotero-mcp was running with pre-edit env vars even after the gateway was killed and restarted).