Skip to content

Checklist

Go through this list before you share a Connector. Each item is covered in more detail elsewhere in this section, and most of them are things that fail silently or confuse the model if they are missed.

Names

  • [ ] The Connector's name gives a sensible slug, and {connector slug}__{tool slug} is under 64 characters for every tool.
  • [ ] Tool slugs are short snake_case verbs, such as search_issues.
  • [ ] Input names and auth profile names are valid Python names, and not keywords.
  • [ ] Every input that names a file or folder has its Path set.
  • [ ] No tool's Uses auth row shows a name in red.

Where It Runs And How It Signs In

  • [ ] Tools that work with the person's computer are in a Client — connected desktop Connector; tools that call web APIs are in a Server — gateway runtime Connector.
  • [ ] A Client Connector uses only Static credential profiles.
  • [ ] Credentials come from auth.<name>, never from an input, a default value or the code.
  • [ ] Each auth profile's description says where to get the credential.

Scripts

  • [ ] result is set at the top level, on every path — including errors.
  • [ ] result holds only JSON types, and a typical result is under about 11,000 characters, with hard limits on lists and long text.
  • [ ] Expected failures return {"error": "<what happened, and what to do next>"}.
  • [ ] Optional inputs cope with None, and numbers are clamped.
  • [ ] No sys.exit(), no reading from standard input, no subprocess that might wait for input, and no relative paths.
  • [ ] Every network call and subprocess has its own timeout, shorter than the tool's.
  • [ ] The tool's Timeout (seconds) is set.
  • [ ] Only the standard library and the bundled packages are used — or the extra packages are listed under Python dependencies, and the description says why.
  • [ ] Any environment variable a tool needs, other than the minimal set, is listed under Environment variables.
  • [ ] A tool that returns images also describes them in text.
  • [ ] Each tool has been run in the debug runner with realistic inputs, including a failure, and tried in a chat.

Descriptions

  • [ ] Every tool's AI Description says what it returns, when to use it, and which other tool to use instead.
  • [ ] Every input's description gives its format, its default and its limit.
  • [ ] The Connector's description tells an administrator what to set up: credentials, Constant input values, and Python packages.
  • [ ] The Version has been raised since the last file you shared.

What A Connector Cannot Do

Some things are outside what a Connector can do. If what you need is on this list, look for the nearest design that works.

A Connector cannot Instead
Remember anything between calls Every call is a fresh process. A Client tool can keep a file somewhere you choose; a Server tool should rely on the system it talks to
Act as the person using it Credentials are one per installation, and a tool does not know who called it
Report progress or ask a question while it runs Return a result, and let the model ask the person
Run on a schedule, or react to events Tools run only when a model calls them. Use a Scheduled Task that calls the tool
Return a file Return text, a path (Client Connectors), or a link. Pictures can be returned — see Returning images
Run in web chat or on the API, if it is a Client Connector Make it a Server Connector, if what it does can run on the Gateway
Run code other than Python directly A Client tool can run a program installed on the person's computer with subprocess. Check for it with shutil.which first, and fail with a clear error if it is missing