Agent tools

The agentic-routines mcp command is a Model Context Protocol server that your agent starts over stdio. setup writes the config for each agent so you never run it yourself.

Browser tools

Your agent works in its own tab unless told otherwise. Everything it does with these is recorded.

Tool	What it does
`tabs`	List open tabs: id, title and address.
`navigate`	Open an address in the agent tab. Lists saved routines for the site, if any.
`read_page`	The page as an outline of roles and names. Interactive elements get a ref such as `ref_12`.
`click`	Click an element by ref.
`type`	Type into a field by ref. Replaces what is there. `submit` presses Enter after.
`select`	Pick an option in a select, by label or value.
`upload`	Attach a file from this machine to a file input.
`press`	Press a key in the focused element: Enter, Escape, Tab, ArrowDown.
`screenshot`	The area around the last element, or the whole tab with `full: true`. `read_page` costs less.

Routine tools

Tool	What it does
`rt_<name>`	One tool per saved routine, with its parameters. Replays without a model.
`routine_run`	The same, by name, for clients that read tools only at startup.
`routine_resume`	Continue a paused replay. `approve` goes past a pause or stops. `choice` picks the element for a changed step.
`routines_list`	Saved routines and what each is for.
`routine_describe`	The full notes: when to use it, when not to, values it asks for, where it pauses.
`run_steps`	The numbered steps of a recorded run, to choose what to leave out when saving.
`routine_save`	Save a recorded run as a routine. Write the description as what it does, then "Use when ...".

At most 30 routine tools are exposed at once. The rest are reached with routines_list and routine_run.

Directory tools

Tool	What it does
`routines_search`	Search the public directory by words, or by the site a person is on. No account needed.
`routine_install`	Install `owner/name` from the directory. It is checked again on this side.
`routine_install_from_git`	Ask to install from a GitHub repository. The side panel shows every step, and Install is the person's click.
`routine_publish`	Ask to send a routine for review. The person approves in the side panel. An agent can't publish by itself.
`routine_submissions`	Where each submission stands: waiting, rejected with the guideline, live or turned off. Read only.

A replay that stops

A replay returns one of these:

Done. The result text lists the steps run and the time.
Paused. The next step can't be undone. Call routine_resume with approve: true to go on, or false to stop.
Needs a choice. A step changed. The result shows the recorded and current crops as text and up to three candidate elements. Call routine_resume with the ref of the right one.