Your AI provider

Stump drafts with your company's own AI provider and model, with your key. Nothing is drafted through anyone else's account.

Set your provider

An owner or admin opens Settings › Connections › AI provider and enters:

• Provider address: any OpenAI-compatible API, such as https://api.openai.com/v1, https://openrouter.ai/api/v1, a Cloudflare AI Gateway address, or, when Stump runs on your own computer (rer dev), a server on that computer (http://127.0.0.1:PORT/v1). It must be https on the public web, or plain http on this computer; a deployed Stump reaches only public addresses.
• Model: the model's exact id, such as gpt-4.1-mini.
• Key: the provider's API key. Stump seals it and never shows it again; the screen shows its last four characters.

The same setting works locally and once deployed, and an agent can read or change it through stump.setup.ai (reading it is standard work; changing it is a security change under your company's policy). Remove takes the provider out; drafts from a source then stop until a new one is set.

A model on your own computer

Any OpenAI-compatible server on this computer works, with no account anywhere. With Ollama:

1. Install it and fetch a model: ollama pull llama3.2.
2. In Settings › Connections › AI provider, enter the address http://127.0.0.1:11434/v1, the model llama3.2 and, as the key, any word such as ollama: a server on your own computer ignores it, but the field takes one.
3. In Compose, paste a changelog link under From a source, pick the accounts and choose Draft from source.

LM Studio's server works the same way at http://127.0.0.1:1234/v1. A small local model writes plainer drafts than a hosted one; every draft still goes through the claim check.

Through the installation's settings instead

An installation can also set the provider before anyone signs in, through Base's AI settings:

Locally, put them in the checkout's private .dev.vars file and start rer dev again. For a deployed Stump, set them as secrets in the environment you deploy from; rer deploy uploads them (rer deploy --dry-run lists them as optional). A provider saved in Settings is used instead of these, and Settings says which one applies.

rer dev --models live --models-service URL is a different path: it links a capped dev app on a Just Every Models service for local testing, with the key piped on stdin. It is not needed for your own provider.

What a request sends

Only what the draft needs: the source, the person's voice guide, the claims and the venue's rules.

Cost

Each request records its model and its cost in US dollars on the draft.

When the provider fails

With no provider set, Compose says so as soon as you put a link under From a source with no text of your own, with a link to Settings › Connections, before you press anything. A post you write yourself needs no provider. An agent or a hand-off that asks for an AI draft while no provider is set is refused at once with the same reason, before any approval is asked for. When a request fails, the draft is not written, and Compose shows why: no provider set yet, a key the provider refused, a limit or an outage, with the provider's own message where it gave one. Stump never fills in generic copy instead. A draft that fails the claim check or the venue check stays a draft, with its flags.

Where AI text is banned

A venue whose rules ban AI-written text gets a brief and facts, and no request is made to the provider. See Hand-offs.