AI Providers¶

An AI Provider is a model endpoint you supply the credential for. Configure one and OpsPilot's own AI panel comes to life: terminal analysis, error diagnosis and the chat panel all run through the provider you made active. Ten providers ship built in, and one of them never leaves your machine.

The AI Providers catalogue listing ten supported providers

The AI Providers page. Ten providers ship; the credential fields differ per provider, and Test verifies them before you save.

Configuring a provider¶

Open Settings → AI Providers. Each provider row expands to its own fields. Enter the credential, pick a model, click Test, then Save. The active provider is used for terminal analysis, error diagnosis and the AI chat panel.

Test checks the credential against the provider's own API and surfaces the provider's error message verbatim, so an invalid key or a model you do not have access to fails there rather than in the middle of an incident.

You can configure several providers and switch per session from the provider dropdown at the top of the AI panel — a fast local model for routine questions, a larger cloud model for hard ones. The choice is per session, so two tabs open against two hosts can use two different providers at the same time.

The ten providers¶

Provider	Notes
Anthropic Claude	API key
OpenAI / ChatGPT	API key, custom base URL supported
GitHub Copilot	Sign in with GitHub, no API key
Google Gemini	Sign in with Google, or an AI Studio key
Groq	Very low latency inference
Mistral AI	API key
Ollama (Local)	Runs on your machine. Nothing leaves it.
Azure OpenAI	Sign in with Microsoft, or an Azure key; your own tenancy and deployment
OpenRouter	One key, 100+ models, free tiers for evaluation
Custom / Self-hosted	Any OpenAI-compatible endpoint, custom model name allowed

Every provider's exact credential fields, base URL, authentication flow and model list are tabulated in the AI provider matrix. The notes below cover the providers with something to set up beyond a key.

GitHub Copilot¶

No API key. OpsPilot uses a device authorisation grant (RFC 8628): it shows you a code, you approve it in a browser against your GitHub account, and the access token is filled into the credential field for you. Requests go to GitHub's fixed Copilot API endpoint — there is no base URL to set.

Useful if your organisation already licences Copilot and would rather not procure a second AI budget line.

Azure OpenAI¶

Also a device authorisation grant, through the Microsoft identity platform, so it works against any Azure AD tenant. You can paste an Azure API key instead if you prefer.

Azure OpenAI is the one provider where the base URL carries real information: it is an endpoint plus deployment URL, of the shape https://RESOURCE.openai.azure.com/openai/deployments/DEPLOYMENT. The deployment name is part of the path, not a separate field, so a new deployment means a new URL. See Self-hosted & enterprise for why this provider matters to regulated buyers.

Google Gemini¶

Two ways in. Sign in with Google runs a full OAuth 2.0 authorisation code flow with PKCE — a browser opens, you click Allow, and that is the whole setup. Or paste an AI Studio API key into the API Key field if you would rather not use the sign-in.

Gemini is reached through its OpenAI-compatible interface, so it behaves like every other chat-completions provider.

OpenRouter¶

One key, a large catalogue, and free-tier models marked free in the picker, which makes it the cheapest way to evaluate OpsPilot end to end. The free lineup changes often, so OpsPilot ships a static fallback list and refreshes the live catalogue of free models when it can; if the refresh fails you keep the cached list.

OpenRouter's model catalogue is public and answers successfully for any key, including an invalid one, so Test checks the key against an endpoint that requires authentication instead.

Custom / Self-hosted¶

Accepts any OpenAI-compatible endpoint. Two fields: a Base URL and an optional API Key — leave the key blank if your endpoint does not need one. This is the only provider that allows a custom model name, because OpsPilot cannot know what you have loaded; the catalogue ships no model list for it.

Use this for vLLM, LM Studio, a llama.cpp server, LocalAI, or an internal AI gateway. See Self-hosted & enterprise.

What configuring a provider does not change¶

The model still cannot execute anything. Its reply is a proposal, and a proposal becomes a command only when you approve it. No provider, model or setting changes that.
Redaction still runs first, using the Data Handling profile resolved for the connection you asked about.
AI access is still scoped per connection. Enable AI on the connection is what admits a session to the AI panel at all, and a connection with it off is invisible to every provider you have configured. Set it deliberately per connection and check it before you connect.

A cloud provider is not an air-gapped deployment

Configuring any provider other than Ollama or a self-hosted endpoint means your workstation reaches that provider over the internet. Your servers still do not, which is what most regulated teams need — but call it private or VPN-only, not air-gapped.