Nous Research's open-source AI assistant, Hermes Agent, now supports Codex as an application server runtime. When activated, all terminal commands, file edits, sandbox operations, and MCP tool calls for openai/ and openai-codex/ models are executed inside Codex. Hermes steps back to become a shell, handling only session databases, slash commands, gateways, memory, and skill reviews.
Hermes's approach mirrors OpenClaw's earlier strategy, but goes further with tool callback mechanisms. Codex's native tools only cover terminal and file operations. Hermes registers itself as an MCP server for Codex, injecting 17 tools — web search, browser automation, visual analysis, image generation, etc. — via stdio callbacks into Codex's toolchain. When a model calls web_search inside Codex, the request is forwarded via the MCP protocol back to Hermes, and results are fed back to Codex.
There are trade-offs. Four tools — delegate_task, memory, session_search, and todo — rely on Hermes's runtime state and cannot be driven by stateless MCP callbacks, so they are unavailable when switching to Codex runtime. Also, Hermes and Codex have completely separate OAuth authentication; users must log in to each separately to avoid token refresh conflicts.
This feature is opt-in. Default behavior remains unchanged. Use "/codex-runtime codex" to enable, and "/codex-runtime auto" to switch back.