Neural Inverse has no hosted default model. Chat and agents use the models you connect: your organisation's models through Govern, a cloud provider with your own key, or a model server on your own machine or network. You connect them on the Models page, which the Editor and the Agents app share.
Open the Models page
Run Manage Models from the Command Palette, or choose it from the Manage (gear) menu or the Preferences menu, next to Settings.
The page has these parts:
| Part | What it does |
|---|---|
| Your models | Every model of every connected provider in one list, with search. Add a model or deployment id a provider does not list, or remove one from the list. |
| Found on this computer | Providers the Editor found and can connect in one click (below). |
| Subscriptions | A plan you already pay for, where you sign in on the provider's own page. Today: Sign in to ChatGPT, used by the Codex engine. |
| Provider blocks | One block per provider type, with its fields, Connect, Test, Refresh models, Remove and Add another. |
| Default models | The default model for each feature. |
Connect your organisation's models
The provider Organization models (Keystone) calls your organisation's Govern model gateway (<api_url>/api/v1/models/gateway/v1, OpenAI-compatible) with your own sign-in. It has no fields: the address is the signed-in session's verified deployment and the credential is your access token, so neither a settings file nor an agent can point it elsewhere. Your organisation's gateway decides which models you may use, and applies its budgets, data loss prevention, guardrails and request logs.
When a chat request used organisation context, the call carries the id of that retrieval so Govern can join the two in its evidence. If the gateway cannot join them, chat shows a notice; the answer itself is unaffected.
See Govern for how administrators assign models.
Connect a cloud provider
- In the provider's block, fill in the fields. Required fields are marked; Get a key opens the provider's key page where there is one.
- Choose Test to check the key and list the models ("Works, N models available").
- Choose Connect. The block shows Connected., or "Saved, but it could not be reached" with the reason.
| Provider | Fields |
|---|---|
| Anthropic, OpenAI, Google Gemini, OpenRouter, Groq, xAI, Mistral, DeepSeek, Fireworks AI, Cerebras, Qwen (Alibaba Cloud), Moonshot (Kimi) | API key |
| GitHub Models | GitHub token with models:read |
| Azure OpenAI | Resource name, API key |
| Google Vertex AI | Project ID, Region, Service account key (JSON) |
| AWS Bedrock | Region, and a Bedrock API key or access key ID, secret access key and optional session token |
| LiteLLM | Endpoint (default http://localhost:4000/v1), optional API key |
| OpenAI-compatible endpoint | Endpoint, optional API key, optional extra headers (JSON) |
Vertex AI's OpenAI-compatible endpoint cannot list its own models, so the page offers a fixed list; Bedrock lists its models and the page adds cross-region inference profiles. You can add any other model id under Your models.
A stored secret shows as "Stored. Leave empty to keep it." Choose Remove the stored value to delete it.
Connect a local model server
These run on your own machine or network and need no key:
| Provider | Default address |
|---|---|
| Ollama | http://localhost:11434 |
| LM Studio | http://localhost:1234/v1 |
| vLLM | http://localhost:8000/v1 (optional API key) |
The page probes these default addresses and lists a running server under Found on this computer, for example "Ollama is running on localhost:11434 (3 models)". Choose Connect. For a server elsewhere, enter its address in the provider's block.
It also lists a provider whose key is already in the environment the app started with, for example ANTHROPIC_API_KEY, OPENAI_API_KEY, GEMINI_API_KEY or GOOGLE_API_KEY, OPENROUTER_API_KEY, GROQ_API_KEY, XAI_API_KEY, MISTRAL_API_KEY, DEEPSEEK_API_KEY, FIREWORKS_API_KEY, CEREBRAS_API_KEY or AWS_BEARER_TOKEN_BEDROCK. Connecting copies the key into the credential store.
Installing and pulling models is done with the server's own tools (for example ollama pull); the Editor does not install them.
For an air-gapped installation, use your organisation's models through a self-hosted Govern deployment, a local server, or an OpenAI-compatible endpoint inside your network.
Choose default models
Under Default models, choose a model for each feature, or No default:
| Feature | Used by |
|---|---|
| Agents | Agent sessions in the Agents app and the Editor |
| Chat | Chat, inline chat and terminal chat |
| Firmware | Firmware assistance |
| Modernisation | Legacy code modernisation (not loaded in the current release) |
| Checks | Compliance and code checks |
Removing a provider deletes its stored keys from this computer, and features that used its models go back to no default. This applies to both apps.
Where keys are kept
Keys are kept in the operating system's credential store and shared by the Editor and the Agents app. The models service runs in the agent host, not in a window. Without a usable credential store, keys are kept only until the app closes, and the page says so. On Linux, install and unlock a Secret Service keyring, for example GNOME Keyring.
What your organisation controls
- Which models you may pick. Picking a model is a governed action. Your organisation's policy can allow, deny or require approval for model globs (
engineRules.models), and an organisation policy delivered by Govern can deny or require approval for models. A denied model is refused when an agent session starts or calls it. - Organisation models. Govern decides which models its gateway serves you and enforces budgets and data rules there. That is the binding control: a modified client cannot get around it.
- Personal keys. A model reached with your own key does not pass through Govern's gateway, so the gateway's controls do not apply to it. Your organisation decides through its policy on the device which models may be used.
Every change on the Models page (connecting, removing, choosing a default) is recorded in the activity record, without any key.
Not used in this release
The repository also contains an older model management service (Ollama library browsing and auto-setup for Ollama, vLLM and LM Studio). Neither app loads it; the Models page replaces it.