Bring your own model key

The AI workspace talks to any endpoint that implements OpenAI’s /chat/completions with streaming. That covers OpenAI itself, OpenRouter, a local Ollama, a vLLM server on your network, and most things that advertise an OpenAI-compatible API.

The published price list says work through your own key spends no credits, because it never reaches our servers. Read One credit is still counted below before you rely on that in the web workspace today.

Adding a provider

Open Settings in the AI workspace and paste a JSON list into Providers:

[
  { "id": "openai", "displayName": "OpenAI", "baseUrl": "https://api.openai.com/v1", "apiKey": "sk-…",
    "models": [ { "id": "gpt-4o-mini", "label": "GPT-4o mini" } ] },
  { "id": "ollama", "displayName": "Ollama (local)", "baseUrl": "http://localhost:11434/v1",
    "models": [ { "id": "llama3.1", "label": "Llama 3.1" } ] }
]

Each entry needs id, baseUrl and a models array; apiKey is omitted for a local server that does not want one. Save, and the models appear in the picker grouped under displayName — see Choose a model.

An entry that is missing a required field is dropped rather than half-loaded, and saving a list that parses to nothing tells you so instead of silently emptying your settings.

Where the key goes

Into this browser’s local storage, and out again only in the Authorization header of the requests your browser makes to the baseUrl you typed. It is not sent to Tincograph, not stored on your account, and not included in your data export.

Two consequences worth planning around:

• It is per browser. The provider list does not sync. Add it again on each browser, profile or machine you use.
• Anything with access to that browser profile has the key. Use a key scoped to what you need, and revoke it at the provider if a machine goes missing. Clearing the workspace’s local data removes it, as does Clear local data in Settings.

The built-in provider

A mock provider called Tincograph is always present and is the default. It runs entirely in the browser and replies by echoing what you typed — it is there so the workspace works with no keys at all, not because it is a model. Add a real provider before judging the answers.

Your own key never counts

Messages to a provider you added go straight from your browser to that provider, with your key. They never touch our servers, so they never count toward your weekly usage limit. Only replies from Tincograph’s own models do, and those are metered on our side (docs: usage limits).

Signed out, nothing is counted at all — the workspace runs with no account, and credits are an account concept.

Until the workspace stops counting messages it does not serve, the two honest ways to use your own key without spending an allowance are to use it signed out, or to watch the credits figure next to the composer and treat it as counting messages rather than work.

Rate limits and outages are the provider’s

When you use your own endpoint, its availability, its rate limits and its bill are between you and that provider. Our status page cannot report on it.

Credits are only involved when you use ours instead: What a credit is.