Announcement · May 15, 2026 · 4 min read
Introducing Tools in Company Agents.
Agents without Tools are just talk. Tools are how Agents do real things: send messages, write code, update databases, hit your API.
Tools are how Agents touch the outside world. A Tool can be an API call, a file read, a database write, a third-party integration, or a custom function. Without Tools, an Agent is a model with a prompt. With Tools, it does real work.
Built on MCP
Tools in Company Agents are built on the Model Context Protocol. That means a Tool written for one Agent works for any Agent that has the matching Skill. It also means Tools from outside Company Agents (anyone publishing MCP servers) can be added with no glue code.
What kinds of Tools
A few categories. Read tools let Agents see files, search the web, query a database, read an email inbox. Write tools let Agents send messages, commit code, update records, schedule meetings. Integration tools connect to specific services: GitHub, Slack, Linear, Stripe, your internal APIs. Custom tools are anything you write.
Read tools are cheap to give. Write tools need a Skill to scope them, and often an approval gate in the Workflow.
The rough rule
Allowlists, not lockdowns
Every Agent has an allowlist of Tools it can call. The allowlist is set by the Skill the Agent loads. This means the same Agent, on the same Issue, can have very different reach depending on which Skill is active. A research Skill might allow web search and document reading. A deployment Skill might allow database writes and git pushes.
What ships today
Tools, MCP support, and a starter set of integrations are live. New Tools and integrations land regularly. You can also write your own as an MCP server and load it as a custom Tool.