Keep your own backend

Install only the UI item:

pnpm dlx shadcn@latest add https://glossen.vercel.app/r/glossen.json

If pnpm stops with ERR_PNPM_IGNORED_BUILDS, see Troubleshooting.

Then point the provider at your endpoint:

<GlossenProvider endpoint="/api/assistant">{children}</GlossenProvider>

What your endpoint must doCopy Anchor Link

Glossen uses the AI SDK's DefaultChatTransport. Your endpoint receives a POST with { id, messages, trigger, messageId } and must answer with an AI SDK UI message stream (for example createUIMessageStreamResponse).

Plain text chat is not enough. Every user message can contain two custom data parts:

• data-page — the page the reader asked from: { title, url }
• data-passage — a selected passage: { title, url, text, heading? }

The AI SDK drops data parts when it converts UI messages for the model, unless you convert them yourself. Without this, the model never sees the reader's page or selected passages. See the message contract for the full schema and an example convertDataPart.

Optional: verified sourcesCopy Anchor Link

The sources list under each answer comes from a data-sources part: [{ title, url }]. Glossen's backend builds it by checking every documentation link in the answer against the real page list. If your endpoint doesn't send it, answers still work; they just have no sources list.

Optional: output-limit labelCopy Anchor Link

If your endpoint sets message metadata { finishReason: "length" } on the finish chunk, Glossen tells readers the answer stopped at the output limit.

ErrorsCopy Anchor Link

When your endpoint returns a non-2xx response before streaming, Glossen shows the response text as the error message. Return short, safe messages rather than raw provider errors.