Building tools
When the tool your agent needs does not exist, there are three ways to make it.
Ask the AI to build it
An agent with the actions:build scope calls
create_action with a plain-language description, for example "post a
daily summary to #ops". ToolGlue places a refundable hold of $0.50 on your
balance (the amount is always stated before anything is spent), then an AI build loop
researches the API, drafts the tool, and tests it against one of your live
connections. Only connections you have flagged as allowing build testing are used.
The loop fixes and retests up to eight times before giving up.
The hold is not the price. When the build finishes it settles at what the models
actually used, never more than the hold, and the unused remainder refunds to your
balance automatically: a $0.50 hold on a build that used a few cents
settles at those few cents. A failed build refunds what was not spent. Watch
progress live on the build page, including per-step token counts and cost against
the hold. Rebuilds work the same way through update_action or the
"rebuild with feedback" box on the tool's detail page.
Agent-initiated builds also obey your workspace's tool policy: by default an agent needs an open tool window (Settings, "Agent tool changes") before it may create or rebuild anything. See the Security article for how the window works.
Brief the builder for a one-shot build
The single biggest lever on build quality is context. The build dialog (and the
context argument on create_action) takes anything you or
your agent already know: docs excerpts, the exact endpoint and its request and
response shapes, auth quirks, a sample response. The builder treats that context as
authoritative, plans from it, and judges the test run against it, so a well-briefed
build usually lands on the first try instead of paying for trial and error. An
example input helps too: pass one and the builder shapes the input schema around it
and uses it for the live test. The connector's docs page is also fetched and read
automatically when it has one, and deleting a tool later is a soft delete you can
undo from the tool's page.
Write it by hand
The manual editor at Actions, New action creates a tool without any AI
involvement, for free. You choose a key, a target connector, a JSON Schema for the
inputs, and an implementation: either an API template (method, path, body with
{{input.x}} placeholders, and an optional JSONPath to pick the output)
or JavaScript code that runs in the sandbox. Manual saves are validated server-side
and go live immediately, skipping the AI build loop.
Test it in the Playground
The Playground page is a test bench for every tool in the workspace: saved actions, external server tools, and an ad-hoc raw request bench. It generates a form from the tool's input schema, shows the real per-run price above the Run button, and dispatches through exactly the same pipeline agents use, so billing, security checks, and run recording are identical. Test there before pointing agents at a new tool.
Bring a prebuilt server
Existing MCP servers from npm or PyPI can join the workspace too; see External MCP servers.