Custom tools
Build a tool once and every agent can call it: a script, a public web API, or a prompt template. Agents can build them too. You can switch one off, read its code or delete it.
Updated 2026-10-13
Custom tools live in the Team tab, next to the built-in ones. They are shared by the whole workspace and called as tool_<name>.
Three kinds
| Kind | You write | Runs |
|---|---|---|
| Prompt | A prompt with {argument} placeholders, such as “Score these headlines for {audience}: {headlines}”. | As one AI call (counted in credits). |
| Web API | A public https URL with {argument} placeholders, such as https://hn.algolia.com/api/v1/search?query={query}. GET or POST. | From the server, to public hosts only. No headers, so no secret rides on it. A POST asks for your approval and autopilot never skips it. |
| Script | Python, Node or bash. Read arguments from the JSON file at $PYTHE_ARGS (or env ARG_<NAME>) and print the result. | On the calling agent’s own computer. Refused if it tries to send or publish. See Guardrails. |
Make one
- Open Team, then Tools. Choose Add a tool, or add a ready-made one.
- Name it. Lower case letters, digits and underscores, starting with a letter, such as
headline_scorer. - Say what it does. One line, including when an agent should use it. Agents read this to decide.
- Choose a kind and write the body. A hint under the box shows the right format.
- Add its arguments. Each has a name, a type (string, number or boolean), a description and whether it is required.
- Save. It is ready for the whole team at once.
Ready-made tools
| Tool | Kind | What it does |
|---|---|---|
headline_scorer | Prompt | Scores headlines or subject lines 1 to 10 for clarity, specificity and curiosity, with a rewrite for each. |
hn_search | Web API | Searches Hacker News stories to see what builders say about a market or competitor. |
readability_check | Script | Words, sentences, average sentence length, long sentences and a Flesch reading-ease score. |
Managing them
- Each tool shows how many times it has been used, when, and the last error.
- Switch a tool off without deleting it.
- Read its code or prompt before trusting it.
- Delete it. Up to 30 tools per workspace.
- Agents can build tools themselves with
create_tooland list them withlist_tools. You can see who made each.