Everything about provider configuration happens inside VS Code. There is no browser step, no config file to hand-edit, and no API key that has to be pasted into a terminal.

Open it with GitPilot: Settings from the command palette (Ctrl/Cmd+Shift+P), then choose AI Providers.

The first thing you see is a summary, not a form:

AI Providers

  ● GitPilot Server   Connected
    http://127.0.0.1:8000                    [Reconnect] [Change server]

  ACTIVE PROVIDER
    Claude (Anthropic)                            Active  ›
    claude-sonnet-4-5

  AVAILABLE PROVIDERS
    OllaBridge Cloud                                      ›
    Free hosted models — sign in, no API key needed
    Ollama (Local)                                        ›
    Run models locally on your computer
    ...

Click any provider to open its configuration page. Only one provider's form is ever on screen. ‹ Back to AI Providers returns to the overview without leaving the settings tab.

A provider is only activated when you press Save and activate. Opening a page and closing it changes nothing.

Keys are stored by the GitPilot backend, never in VS Code settings and never in your workspace.

The settings page never receives a stored key. It is told only that one exists, and shown the last four characters:

API key configured: ••••A7X2. Leave the field empty to keep it.  Remove API key

Three rules follow from that:

• An empty key field means "keep the current key." You can change a model or a base URL without re-entering the secret.
• Removing a key is a separate, confirmed action — the Remove API key link, which asks before clearing.
• Keys never appear in logs, notifications or error messages. Error text is redacted before it is displayed.

If your GitPilot server is remote and reached over plain HTTP, GitPilot warns before sending a key and lets you cancel.

Free hosted models, and the default. Three connection methods, shown as tabs:

Cloud Login — click Sign in with browser. Your browser opens the OllaBridge sign-in page and shows a pairing code; paste that code back into VS Code and press Pair device. GitPilot never asks for your password. Once paired the page shows your connection and a Sign out action.

API Key — for an OllaBridge endpoint you already have a token for. Set the endpoint, the key, and a model.

Local Gateway — for a self-hosted OllaBridge. Defaults to http://127.0.0.1:11435.

!!! warning "Not port 8000" http://localhost:8000 is GitPilot's own backend, not an OllaBridge gateway. Pointing OllaBridge there makes model discovery query GitPilot about itself. GitPilot rejects that URL with an explanation, and repairs the value automatically if an older install stored it.

Set the Ollama URL — http://127.0.0.1:11434 by default — and pick from the models that machine has pulled. Refresh re-scans.

If Ollama is not running, the page says so plainly and offers Retry and Install Ollama rather than leaving an empty dropdown.

Enter the instance root, not the API root. GitPilot talks to Ollama's OpenAI-compatible surface and appends /v1 itself; typing it yourself is harmless — the URL is normalised, so http://127.0.0.1:11434 and http://127.0.0.1:11434/v1 mean the same thing and neither becomes /v1/v1. Model ids are passed through exactly as Ollama reports them, so anything ollama list shows works, including ollama create names and hf.co/user/model pulls.

Needs an Anthropic API key from console.anthropic.com.

!!! note A Claude.ai subscription does not include API access. The API is billed separately and needs its own key.

Model and an optional custom base URL are configured on the same page.

An API key from platform.openai.com, a model, and an optional base URL — set the base URL for Azure OpenAI or a proxy.

Needs both an API key and a project ID; watsonx is not considered configured with only one. The region base URL defaults to https://us-south.ml.cloud.ibm.com.

Point GitPilot at your Open WebUI instance — http://localhost:3000 by default. Paste the root URL: GitPilot appends the OpenAI-compatible API path itself, so /api or /v1 on the end is unnecessary (and handled if you include it anyway).

The API key is optional — an instance open to your local network needs none. If yours requires one, create it in Open WebUI under Settings → Account → API keys.

Models are discovered from the instance. If it does not answer, the page says so instead of showing an empty list.

Any OpenAI-compatible chat-completions gateway: self-hosted inference, a corporate model gateway, a proxy in front of several vendors.

Request headers exist because gateways routinely require attribution or routing headers alongside the key — for example a header carrying your user identity so usage is attributed correctly:

x-user            you@example.com
x-client-app-id   gitpilot

Add rows with Add header, remove them with Remove. The saved set is exactly what the editor shows: a row you delete is deleted, and a row with a blank name is ignored. Put the token in the API key field, not in a header.

Model discovery is best-effort and tries the richer source first:

1. A published catalogue at the endpoint's origin (/.well-known/opencode or /.well-known/models)
2. The OpenAI-compatible /models listing

If neither is available — plenty of gateways serve chat-completions without a catalogue — the page says so and you enter the model id by hand. That is a working configuration, not an error.

The settings page always opens. If the backend is unreachable you get recovery actions rather than a dead-end dialog:

GitPilot server is not connected

Provider settings are stored by the GitPilot backend, so they
cannot be loaded right now.

[Start local server]  [Reconnect]  [Change server URL]  [Copy diagnostics]

Or start it yourself:
gitpilot serve --no-open

Start local server runs gitpilot serve --no-open for you and follows the port it actually binds — GitPilot moves to the next free port when 8000 is taken, and the extension follows it rather than losing the connection. Output goes to the GitPilot output channel.

For a remote server URL there is no Start local server button: starting a process on your machine would not make someone else's server reachable. You get Reconnect, Change server URL, and Copy diagnostics.

If gitpilot is not on your PATH, set gitpilot.serverCommand to its full path.

Provider pages avoid the slow endpoints deliberately:

/api/status is not called from these pages at all; it probes live providers and can take upwards of 15 seconds.

Provider credentials are deliberately not in this table. They live on the GitPilot server, not in VS Code configuration, so they are never written to settings.json or synced by Settings Sync.

Configuring in VS Code is the recommended path. For headless or scripted installs, the backend also reads:

Environment values are merged in on every load, so operators can supply credentials without them being written to disk. Custom request headers are configured in the settings UI.

See .env.template for the annotated list.

"OLLABRIDGE API key not configured" on startup Fixed. OllaBridge needs no API key, and the startup check no longer claims otherwise. Upgrade if you still see it.

Model list is empty Discovery is lazy — press Refresh. If the provider cannot be reached the page shows why. Every provider page accepts a manually typed model id.

"Not connected" when the server is running Fixed. The connection state is re-probed rather than read from a cached flag that refreshed on a 30-second timer.

Changing the provider dropdown did nothing That dropdown is gone. Provider selection is a page with an explicit Save and activate, which writes through the backend API.

"Fallback to LiteLLM is not available" when a task produces a plan Fixed in 0.2.8, and it affected Ollama, OllaBridge, Open WebUI and custom endpoints alike. Chat reaches these providers over plain HTTP, but planning hands a model to CrewAI, which would only route a provider it recognised natively — and on CrewAI 1.6 and earlier, none of these four qualified. Upgrade GitPilot; if you cannot yet, pip install litellm unblocks it. Full explanation.

• Agent architecture and topologies
• Ports and defaults