Skip to content
AI Primer
release

Nous Research integrates Codex app-server into Hermes Agent for OpenAI tool runs

Hermes Agent can now route core tool calls through the Codex app-server when it is using OpenAI models. The integration gives Hermes users access to Codex runtime behavior with a `hermes update`, without changing the rest of their agent stack.

3 min read
Nous Research integrates Codex app-server into Hermes Agent for OpenAI tool runs
Nous Research integrates Codex app-server into Hermes Agent for OpenAI tool runs

TL;DR

  • NousResearch's launch post says Hermes Agent can now hand openai/* and openai-codex/* turns to the Codex CLI app-server, so Codex becomes the runtime for terminal commands, file edits, sandboxing, and MCP tool calls.
  • The same launch screenshot says the integration is opt-in, Hermes does not auto-route users onto Codex, and default Hermes behavior stays unchanged unless the flag is flipped.
  • In the docs screenshot, Nous frames Hermes as the shell around Codex, keeping the sessions database, slash commands, gateway, memory, and skill review on the Hermes side.
  • Teknium's reply says the feature is available now after hermes update, which makes this a live ship rather than a docs-only preview.

You can jump straight to the docs page, and NousResearch's screenshot is unusually specific about the split of responsibilities: Codex gets the tool loop, while Hermes keeps the session and memory layer. The timing also lands just after NousResearch's Hermes Agent Jam post and a burst of hands-on praise like trycua's repost of andyhennie and Teknium's repost of jumperz, which gives the integration a ready-made audience.

Codex app-server runtime

The core change is narrow and useful. When Hermes is using OpenAI models, it can route core tool execution through Codex instead of running its own tool loop.

According to the launch screenshot, the Codex runtime takes over four things:

  • terminal commands
  • file edits
  • sandboxing
  • MCP tool calls

The same docs screenshot also makes the boundary clear: this is optional, limited to OpenAI model turns, and described as a runtime swap rather than a full product merge.

Hermes shell

The interesting architectural detail is what did not move. In the documentation screenshot, Hermes stays responsible for the sessions DB, slash commands, gateway, memory, and skill review, even when Codex is executing the tools.

That makes the integration look more like a backend tool-runtime substitution than a new agent stack. dkundel's reply read it as “Codex everywhere,” but the docs describe a more specific arrangement: Codex powers the tool loop, Hermes still wraps the session.

Rollout

The rollout instruction is one line long: run hermes update. That came minutes after the initial announcement, via Teknium's reply, and matches the “flip of a switch” framing in NousResearch's post.

The release also lands into a moment when Hermes already had momentum in its own community. trycua's repost of andyhennie said Hermes was easier than OpenClaw and that computer use “actually works,” while Teknium's repost of jumperz claimed a smooth build experience after migrating. Those posts are not about Codex itself, but they do show the integration shipped into an active user base rather than an empty demo.

Further reading

Discussion across the web

Where this story is being discussed, in original context.

On X· 2 threads
Hermes shell1 post
Rollout3 posts
Share on X