Custom tools New in 0.9.0

The tool your team keeps wishing for, built for your agents.

Checking a VAT number, turning a bank export into your bookkeeping format, pulling the week's figures from a system only your firm uses. Describe the tool in your own words. Crew's coding agents write it, its tests and its examples, and Flight Deck runs it for your agents.

Off by default. An admin turns it on in Admin console → Settings → Custom tools. Then Custom tools is under Multi-Agent in the full layout and in the kiosk's Settings menu.

Mine Shared with me Everybody Approvals, for admins

From a sentence to a tested tool.

For people an admin approved for Crew. Nothing goes live until its own tests and every example pass.

  1. 1

    New tool

    A Name, a Title and What it should do, in plain words.

  2. 2

    Check it

    Your own model returns a fuller spec, with a reason for each change, and an Already there? warning when agents can do it already.

  3. 3

    Build it

    Keep Checked or Yours and pick a coding agent. The worker writes the tool, its tests and its examples in the tool's own git repository.

  4. 4

    Goes live if it passes

    When you approve the merge, Flight Deck publishes it in the background. If a test or an example fails, the old version stays live.

The last three versions can be made live again under Versions. Try it on the tool's page runs it by hand, and has Run as for tools that call an agent.

custom_tools A new agent tool

Flight Deck runs it, never the agent.

Agents list, describe and run your tools through custom_tools, and can ask Crew to build or improve one, or propose one for you to build. Every call is one process, JSON in and out, run as the person whose turn it is.

Calls get 60 seconds (300 at most) and 1 GB of memory by default. The tool is shown to agents only while the deck has Custom tools on.

Build it once. Offer it to the team.

One person makes the tool, and the people who need it add it for themselves.

A tool can ask for more.

Some jobs need more than a function. A tool says what it needs in its tool.json, with a reason for each, and the person adding it sees that before they agree.

ask_user Also new

An agent asks you, in Flight Deck.

The same dialog lets any agent ask the person its turn is for one question (text, one choice, several, or yes/no) or a short form of up to 10. It shows in every Flight Deck tab they have open, and the bell says ‹agent› asks you something.

The sandbox Linux 5.13 or later

Held in a sandbox, where the host has one.

On Linux 5.13 or later, each run is held in a Landlock sandbox. The admin's card says which kind of host you have.

macOS, Windows and the stock Raspberry Pi 5 kernel have no Landlock. There, tools run as Flight Deck's own user and can reach what it can. Run tools only in the sandbox stops every tool on a host without one. Installing a tool's own packages at publish time runs outside the sandbox, and approving someone for Crew still trusts them with the server. · Who to approve for Crew

Turning it on.

Custom tools stay off until an admin turns them on. Only people approved for Crew can make them; anyone can run one offered to them.

  • Admin console → Settings → Custom tools turns them on for the deck.
  • The host needs git; uv is optional.
  • Back it up. Tools live in <FD data dir>/external_tools/.
  • Check a Linux kernel with python3 tests/test_flight_deck/test_xtool_sandbox.py, which says whether it has Landlock.

Custom tools and ask_user were tested with stand-in models. Details and what's still open are in the Flight Deck guide.

Custom tools docs · Crew writes the tools · The team checklist

Keep reading

Try it on real work.

Join the waitlist for Cohort 001, our first early-access group, and you'll get an email when you're in. Or run Captain on your own machine today.