Hermes Agent Integrates Codex Runtime: All OpenAI Model Calls Handled by Codex, MCP Tools Callback I

Following openclaw’s lead, Nous Research has introduced a new Codex APPlication server runtime for its open-source AI assistant, Hermes Agent. When enabled, all terminal commands, file edits, sandbox Operations, and MCP tool calls originally routed through OpenAI/* and openai-codex/* models are executed entirely inside Codex. Hermes then ACTs as a Lightweight shell, manAGIng only the session database, Slash Commands, gateway, mEMOry, and Skill review.

While Hermes shares the Same underlying approach as OpenClaw, it extends the tool callback mechanism further. Codex natively supports only terminal and File Operations. Hermes registers itself as an MCP server to Codex, injecting 17 additional tools—including Web Search, browser automation, visual analysis, and Image Generation—into Codex’s toolchAIn via stdio callbacks. For example, when the model calls web_search inside Codex, the request is forwarded to Hermes through the MCP protocol, and the result is projected back to Codex.

There are trade-offs in functionality. Four tools—delegate_task, memory, session_search, and todo—depend on Hermes’ runtime state and cannot be driven by stateless MCP callbacks, making them unavailable under the Codex runtime. Additionally, OAuth authentication remains completely isolated between Hermes and Codex, requiring separate logins to avoid token conflicts.

The feature is opt-in, preserving default behavior. Users can enable it with /codex-runtime codex and switch back to auto mode with /codex-runtime auto.