Serverless functions your AI can call.

Same shape, different surface area. Lambda gives you compute and leaves you the rest: an MCP server in front of it, OAuth for the clients that call it, permissions per caller, secret rotation, and a run record an auditor can read. Then a second stack when the same logic has to answer plain HTTP. Kitted is the function plus that surface. If you already run all of it in your own account, you do not need us.

You can, and the server itself is the easy afternoon. What follows is the part that lasts: somewhere to run it, secret storage and rotation, a sandbox because these tools reach production systems, version history with a rollback that isn't a redeploy, an audit trail before someone asks you for one, and the same work again for every agent platform your company adopts. Kitted is that second list, and you still get the code.

Usually you should, through a tool. Your API is the system underneath, and a Kitted tool calls it. What the endpoint does not carry is a contract shaped for a model, the business rules that decide when a call is allowed, permissions per person and per client, and a record tying the call back to whoever asked for it. Wrapping an API you already own is the normal case here, not the exception.

Behind it. Kitted is an origin server: the place a tool is built, holds its credentials, and records what it reached. A gateway aggregates origins and fronts them for your clients, which is a different job. Point yours at your workspace endpoint and keep your RBAC and your single address. The two are not alternatives, and a team that has the gateway usually still has nowhere to build the tools it serves.

A TypeScript or Python script with a typed input, your logic, and a typed result. Dependencies pinned in package.json or requirements.txt, secrets arriving as environment variables, and every invocation in its own sandbox with hard time and output caps. Your assistant sees a tool it can call; you see a script you can open and diff.

Anything that speaks MCP, over stdio or HTTP: Claude, ChatGPT, Cursor, and the agent platforms your company standardises on. The same tool is also a plain HTTP endpoint, so a cron job or one of your own services calls exactly what the assistant calls, under the same permissions.

With the workspace, not in the code, which matters because the code is downloadable. The manifest declares which secrets a tool needs, the platform injects only those into an isolated per-tool environment, and environments are separated so a preview never holds production credentials. Every run records who asked, which tool ran, and which hosts it reached.

Yes. Every tool is a self-contained package: your logic, its dependencies pinned, and a manifest. Download it and it runs with node script.ts or python main.py plus a few environment variables. Nothing in the code calls back to a proprietary runtime, so the same artifact runs on your laptop, in your CI, or here.