Agentic tools

The AI chat runs as a multi-step agent loop: the model can call tools, read the results, and continue reasoning before it answers. The loop is capped at 5 steps per turn (stepCountIs(5)), and tool results stream back to the client as they resolve.

Built-in tools

The tool registry lives at lib/ai/tools/index.ts. Two tools ship with the AI chat layer itself:

Tool	Source	What it does
`calculator`	`lib/ai/tools/calculator.ts`	Evaluates basic arithmetic (`+ - * / % ^` and parentheses).
`data-fetch`	`lib/ai/tools/data-fetch.ts`	Fetches the text of a public web page to summarize or answer questions about.

Selecting the agentic tool calling feature at scaffold time adds one more:

Tool	Source	What it does
`web_search`	`lib/ai/tools/web-search.ts`	Searches the web through one of three interchangeable backends.

When tools are active

Tools are only passed to the model when both conditions hold:

ENABLE_TOOLS is true — a template-admin switch in lib/ai/config.ts. Set it to false to run plain chat with no tools.
The workspace’s plan allows tools — features.tools in plans.ts (off on Free, on for Pro and Team).

When tools are off, the chat still works — it just answers without calling them.

Web search backend

Pick a backend by setting its API key in .env.local:

Backend	Key
Tavily	`TAVILY_API_KEY`
Exa	`EXA_API_KEY`
Firecrawl	`FIRECRAWL_API_KEY`

The tool source lives at lib/ai/tools/web-search.ts.

Adding your own tool

Create a file under lib/ai/tools/ defining the tool — a name, a Zod parameter schema, and an execute function.
Register it in the tool registry.
The chat agent loop picks it up; the model can now call it.

Keep tools tenant-aware — if a tool touches workspace data, scope it by the active tenant_id just like any other query.