Run Mixlayer models in Codex, Hermes and Pi
Coding agents are where open models get their hardest workout, and that changes what "works with Mixlayer" has to mean. An OpenAI-compatible endpoint isn't enough on its own. The harness has to see the right context window, tool calls have to stream properly, and reasoning settings have to map onto what the model actually supports.
We've tested Mixlayer with three agent harnesses, and each now has a setup guide.
Codex CLI
Codex CLI supports custom model providers through the Responses API. We recommend a separate profile, so your normal Codex config stays untouched. Save a profile at ~/.codex/mixlayer.config.toml that sets Mixlayer as the model provider with wire_api = "responses", then run:
codex --profile mixlayerThe guide uses GLM 5.2, a 256K-context model that suits long coding sessions. Run /status inside Codex to confirm the model and provider.
Getting Codex right took some server-side work. Codex sends tools with type: "namespace" in the Responses API, so we added support for that, along with parallel_tool_calls: false.
Hermes Agent
Hermes Agent talks to any Chat Completions endpoint. Run hermes model, choose Custom endpoint, and enter Mixlayer's base URL, your API key, a model ID and its context length. Hermes saves the settings to ~/.hermes/config.yaml.
Pi
Pi is a coding agent harness, and we wrote a small extension that registers Mixlayer as a provider:
pi install npm:pi-mixlayer
piThen run /login, choose Use an API key, pick Mixlayer, and paste your key. pi --list-models shows what's registered.
Which model to use
For long, multi-step coding work, start with GLM 5.2 or Kimi K2.7 Code. Both have a 256K context window and support tool calls. For a quick, free test, Hermes's guide uses qwen/qwen3.5-4b-free, which is meant for prototyping and short tasks.
The full guides are in the docs: Codex CLI, Hermes and Pi. If you run a harness we haven't covered, tell us which one.