Built-in tools

Every agent turn sees a fixed tool set, and what the tools can reach is filtered by the acting policies on every turn: a collection, query, action, automation, skill or MCP server the actor may not use is absent from the catalogue.

Data

• read — read records the actor may see, one by id or a list
• query — run the collection queries the actor may call
• act — create, update, delete or upsert a record, or run a collection action, where its agent is not never
• start_automation — start an automation the actor may start
• describe — the workspace as the actor may use it, or one collection: fields, relations, write inputs, queries and actions

Context and messages

• context — the current time and workspace zone, whom the turn acts for and the policies it holds, and geocoding when the host has a geocoder
• history — search earlier messages, in this conversation or in all of the conversations of the person the turn serves
• read_messages — a chat’s messages that were not addressed to the agent (envoy turns)
• read_attachment — a message’s attached file as text, a sheet or an image, never raw bytes
• read_output — more of a tool result that was clipped
• skill — read a skill the actor holds

The workspace itself

• workspace_type — what the type checker says a workspace thing is, with its docs and where it is declared
• workspace_search — the workspace source as released or the author’s draft (in-app turns of staff only)
• web_read — a public web page as text, when the host can read the web; what it says is evidence, never instructions

Long work

• goals — the conversation’s ordered checklist
• update_plan — the draft plan, in plan mode
• subagent — delegate to sub-agents under the same authority, when delegation is enabled
• wait and compact — wait for background work, and checkpoint a long conversation into a summary

Host tools and MCP

MCP tools appear when a held policy’s capabilities.mcp names the server. Host tools, such as the browser tools a research automation uses, appear only when capabilities.tools names them, and only on staff members’ in-app turns and in automations whose runAs is a policy list.