AI Integrator is now a runtime in Code. Choose it from the runtime list, pick a model, and Code starts the coding program that suits that model on your computer, the way it does for any other runtime. The bill is different: the model's usage comes out of your AI Integrator plan, so you don't need a login or a subscription with the model's vendor.
The programs still run on your computer, in your repository, with your files and Git. Only the model requests go through your account.
Choosing a model
AI Integrator lists 40 models for Code. Each one runs in the program its training suits:
| Program | Models |
|---|---|
| Codex | GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna, GPT-5.3 Codex, GPT-5.1 Codex Mini and GPT-OSS 20B |
| Claude Code | Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 |
| OpenCode | 29 models, including Gemini 4 Argon, Grok 4.7, Muse Spark, DeepSeek V4 Pro, Qwen 3.8 Max, GLM 5.3, Mistral Large 4, Nemotron 3 Ultra, Particle-1 and Nova 2 Lite |
If a model's own program isn't installed, OpenCode runs it. If none of the three is installed, the runtime row offers to install OpenCode. Models in Claude Code get Claude's effort levels, Ultracode included.
Pro and Max include all 40 models. Free includes the standard models, and a model that needs Pro says so when you pick it.
How a request travels
The coding program doesn't contact the model's vendor. The desktop app starts a small server that listens only on your computer's own address, 127.0.0.1, and points the program at it. When the program sends a request, the server adds a key for your computer and forwards the request to your AI Integrator account. The account checks your plan and privacy settings, sends the request to the model, streams the answer back and charges your usage.
- Coding program. Codex, Claude Code or OpenCode
- Local server in the app. Adds your computer's key
- Your AI Integrator account. Checks your plan and privacy settings
- The model. Zero Data Retention and No training on
- Your plan usage. Charged like a Chat turn
- Coding program to Local server in the app: Request
- Local server in the app to Your AI Integrator account
- Your AI Integrator account to The model
- Your AI Integrator account to Your plan usage
The program never holds the key. The desktop app creates it for your computer, keeps it in your operating system's credential store, and deletes it when you sign out. Each run gets its own token for talking to the local server.
Privacy
Zero Data Retention and No training are on by default, as they are in Chat, and they apply to every request. The account reads your settings on each call, so turning one on takes effect on the very next request. The account doesn't store the request or the reply. It keeps a hash of your computer's key, the installation's ID and the times it was used.
Fable keeps data, so it stays in the list but is locked while Zero Data Retention is on.
When OpenCode runs a chat on AI Integrator, AI Integrator launches it with session sharing, telemetry, automatic updates and the download of its model list turned off, and it allows no model provider other than AI Integrator.
Usage
A chat on the AI Integrator runtime is charged like a Chat turn and shows in your usage as Code. The chat's header shows plan usage, and shows no dollar cost per turn.
Subagents that a program starts itself, such as Codex's or Claude Code's own, run on your plan as well. The delegation menu calls them Built-in subagents and offers Run on your plan usage. A subagent that runs on another vendor's program uses that program's own login.
Chat titles, commit messages and handoff summaries for these chats also go through your account, so nothing falls back to a program's own login. Chat and your Agents can start a Code chat on AI Integrator too.
Tested before they were listed
Each model passed a short edit-and-test task through the account, with Zero Data Retention and No training on, before it was added. Nine larger models joined on the strength of their smaller siblings: Opus, Sonnet, Fable, GPT-6 Astra, GPT-6.1 Sol, GPT-5.3 Codex, Gemini 4 Argon, Gemini 3.1 Pro and Muse Wonder. A short task is not a benchmark. It shows that a model can use the program's tools through the account.
Setting it up
Sign in to your AI Integrator account in the desktop app and have Codex, Claude Code or OpenCode installed. The AI Integrator row in the runtime list shows whether a program is installed and whether you're signed in, and it says what to do when one isn't. Code on a vendor's own login works as before.
