In-product agent

Every namespace ships with an AI assistant baked into the console. Open it from the sidebar Ask AI button or press ⌘K (Ctrl+K on Windows/Linux). It runs on Cloudflare Workers AI with the @cf/moonshotai/kimi-k2.5 model and inherits your namespace role on every action.

What it can do

The agent has tools at two levels.

Namespace-level

Tool	Effect
list_databases	Enumerate every database in the namespace
propose_create_database	Stage a new database (always asks before creating)
list_members	Members and their roles
propose_invite_member	Stage an invitation by email + role
list_tokens	List API tokens (without secrets)
propose_create_token	Stage a new API token
get_usage_summary	Last-N-day request / row counts

Database-scoped (every tool takes a database slug)

Tool	Effect
list_tables	Schema dump for a database
describe_table	Columns, types, constraints
run_select	Execute a single read-only SELECT (capped at 200 rows)
ask	NL → SQL → answer for the database’s data
read_docs	Stored table/column descriptions
list_endpoints	Published endpoints for the database
recent_endpoint_logs	Last invocations from analytics
propose_create_table	Stage a CREATE TABLE
propose_create_endpoint	Stage an endpoint spec
propose_scaffold_crud	Stage list / get / create / update / delete endpoints for a table
propose_generate_docs	Stage auto-docs for selected tables

Anything starting with propose_ is a two-step flow — the first call returns the plan, you see it in the chat, and only after you say something like “yes, do it” does the agent re-call with confirmed: true to actually apply.

Limits

There is no per-tier message cap and no per-token throughput cap. Agent usage drains the same prepaid balance as everything else; when the balance runs out you’ll get a 402 with a top-up hint.

Each turn is bounded to 5 recursive tool calls so the agent can’t loop forever.

How it works

1. The console POSTs to /api/namespaces/:slug/agent/chat with your message.
2. The API worker assembles your last 50 turns of history, the system prompt (which embeds your namespace slug + role), and the tool definitions.
3. runWithTools from @cloudflare/ai-utils drives the model loop on Workers AI. Tool invocations execute as you, with your namespace role.
4. The response streams back as SSE: a leading event: trace frame listing every tool call, then the model’s tokens, then [DONE].
5. The assistant turn is persisted to D1 so reconnecting clients see the same conversation.

History is per (user × namespace). Clear it from the chat panel — that hits DELETE /api/namespaces/:slug/agent/history.

Prompting tips

• Be concrete. “Show me last week’s signups” beats “tell me about my data”.
• Reference databases by slug. With one database the agent picks it automatically; with several, name it.
• Approve propose_* actions explicitly. “Yes, create it” or “go ahead” works. The agent won’t mutate without that.
• Greetings get conversational replies. Saying “hi” won’t trigger any tools.