Adding a Provider

​

1. Write the adapter

export const id = 'myvendor';
export const label = 'My Vendor';
export const defaultModel = 'my-current-model';     // pick a CURRENT id
export const models = ['my-current-model', '…'];
export const signupUrl = 'https://myvendor.example/keys';
export const needsBaseUrl = false;                  // true if the endpoint needs a base URL

export async function* stream({ apiKey, model, baseUrl, system, messages, signal }) {
  // Translate `system` + `messages` (canonical content blocks, incl. images)
  // into the vendor's request, call its streaming API, and YIELD text chunks.
  const res = await fetch(`${baseUrl || 'https://api.myvendor.example'}/v1/chat`, {
    method: 'POST',
    headers: { Authorization: `Bearer ${apiKey}`, 'content-type': 'application/json' },
    body: JSON.stringify({ model, stream: true, /* …translated messages… */ }),
    signal,
  });
  // …parse the stream and `yield` each text delta…
}

// Optional, if the vendor supports it:
// export async function generateImage({ apiKey, model, prompt, size }) { … }

​

What stream must do

• Yield text chunks as they arrive (it’s an async generator).
• Respect the abort signal.
• Translate canonical messages — including tool_use / tool_result blocks and image blocks — to the vendor’s shape. Use the shared helpers in server/providers/multimodal.js, sse-util.js, and limits.js so multimodal and streaming behave like every other provider.

​

2. Register it

​

3. Test it

• Re-run onboarding and pick My Vendor; enter a key (and base URL if needsBaseUrl).
• Run a simple turn (<say> should stream).
• Run a multimodal turn (attach a screenshot) to confirm image handling.
• Confirm a tool turn (e.g. write_file) round-trips — the model must read <tool_results> and continue.

​

Conventions

• Provider parity. Your provider must be a first-class peer — never a second-class adapter. The chat loop, tools, surfaces, and bridge must behave identically.
• Model ids go stale. Don’t hardcode an old id as the default; fetch the vendor’s live docs and pick a current model.
• Caching. If the vendor supports prompt/context caching, wire it through the shared prompt-cache layer where the new harness uses it.
• Vision. If the vendor has a vision model, image-driven features (locate_screen_element, browser/screenshot feedback) work; if it’s text-only, they won’t.

​

Checklist

• Adapter exports { id, label, defaultModel, models, signupUrl, needsBaseUrl, stream }.
• stream yields text deltas, respects signal, handles image blocks.
• Registered in ALL_PROVIDERS (server/providers/index.js).
• Appears in onboarding; simple + multimodal + tool turns all work.