AI Provider
zudo-text uses a single Cloudflare Workers AI model as its sole AI provider, served by the dedicated agent-server Cloudflare Worker rather than a third-party API. It is fully operator-provisioned — no model-provider API key or external account is required from the user. Access is gated by your cloud-sync account: sign in once via Settings → Sync for the writing-app AI surfaces.
Sign-in requirement
The AI assistant frame and inline AI command require an interactive signed-in cloud workspace session. The MCP integration's ask_zudo_agent tool reaches the same service with a workspace-bound PAT carrying agent:invoke, not the interactive session. If you have not signed in, the AI assistant shows a sign-in prompt and inline requests return an auth error.
Sign in via Settings → Sync → Sign in using the same Better Auth account you use on other devices.
Per-day quota
Each user has 100 AI-turn admissions per UTC day. Assistant messages, inline-command invocations, and MCP ask_zudo_agent calls share that pool. A turn is charged atomically when the server admits it, before the model runs, so a later model, tool, or stream failure still consumes the admission.
When the quota is exhausted the server returns a quota_exceeded error. The inline command surface shows a one-line notice; the AI assistant panel shows it in the thread. The quota resets at midnight UTC. See Flue Agent Platform → Quota for the server-side accounting.
Model and backend
The model in use is @cf/, served via Cloudflare Workers AI. There is no user-facing model selector — the backend resolves the model server-side, and both AI surfaces (the assistant and the inline command) address the exact same model id.
The two surfaces are served by different backends within the same Worker, agent-server (zudo-agent-server):
The AI assistant is a durable, multi-turn conversation with its own workspace tools (list/read/search notes, check kanban boards, create notes, propose edits), built on the Flue agent platform — see that page for the full architecture: conversation identity, credential handling, edit safety, and retention.
The inline AI command is a one-shot streaming call with no durable conversation and no tools — it exists purely to transform text in the editor buffer.
See Also
AI Assistant — The chat panel: thread tabs, toolbar, skills reference, workspace access, and sign-in empty state
Flue Agent Platform — The architecture behind the assistant and its tools
Inline AI command — The
@@/Cmd+Jin-editor prompt surfaceMCP integration — Using the same assistant from Claude Code or Codex