Sitelooper

Navigate the current tab to a URL and wait for load.

Hover an element (also dispatches synthetic mouseover/enter for JS-driven menus).

Report a background job started by build_start, check (ready), repair_propose or rerecord: status running|done|failed, the last progress lines, and the final JSON result and exit code when finished. Waits up to wait_s seconds for the job to finish before answering.

List saved flows with their step counts, variables and origin.

Point to where each listed value is shown on the current page, so it can be re-read later.

Click while holding modifier keys (e.g. Shift/Control-click gestures).

Navigate a browser session to a URL (no model involved); starts the session if it is not running. learn and headed apply only when this call starts the session: learn=true records every later do/assert as a reusable flow step (needed before save_flow). Typical authoring: open (learn) → set_var → do/assert per outcome → save_flow → build_start.

Return the session page's URL, title and an accessibility snapshot (no model, no actions).

Press a key or chord (e.g. "Enter", "Escape", "Control+a") on an element or the page.

Read text/value/attribute from ONE element — much cheaper than a full snapshot for spot checks. The target must match exactly one element, or the read fails: a snapshot ref (@e123) always does, and a bare tag like "h1" usually does not. Use read_all to read every match, or what=count to count them. If the value matters beyond this glance — a reference, id, name, or total the task or a later step will use — pass `label` NOW: a labelled value is published under that name automatically; an unlabelled one stays anonymous. what=url reads the current page URL (no target) — the way to report where a …

Like read, but returns a JSON array of the value across EVERY element matching the selector — read a whole list of rows/cells in one call instead of many. what: text (visible text of each), value (input value of each), attr (an attribute of each), or count.

Submit the value renames for this session. Rename ONLY names that mislead or say nothing (selector fragments like h1, role_radiogroup_aria_che, generic value/value_2, or a name that contradicts what the value is). Leave every adequate name alone — churn costs more than it buys.

Apply a proposal staged by repair_propose to its flow file exactly as checked (no browser run).

Start a background job that replays a compiled .flow.ts against the app, repairs drifted steps and stages a checked candidate as a proposal file for review (repair --propose); never rewrites the spec. Returns a job_id: poll job_status. Apply a good proposal with repair_apply.

REQUIRED final call: report the outcome of the instruction. Nothing after this is executed. Keep summary to one short paragraph.

Start a background job that re-records one step of a flow with the agent (optionally with a new instruction), then verifies it over several runs and re-pins it. Needs a model API key. Credentials: write each secret as one {{env:NAME}} marker ({{totp:NAME}} for a one-time code), exactly as written: the marker alone, never the value and never $NAME. NAME must be set in the environment the MCP server was started with. Never ask the user for a secret value. Returns a job_id: poll job_status.

Right-click (context menu) an element.

Description is built at runtime.

Replay a stored procedure listed under [skills] in the instruction, deterministically and without further reasoning: every recorded step runs in order with its parameters filled in, stopping at the first step that no longer works. Returns each step's outcome and every value read back from the live page. If it stops part-way, the steps that ran HAVE changed the page — observe, then continue from there yourself. Call it as your FIRST action when a listed procedure matches the instruction.

Stop a learn session, save what it recorded as a named flow, then export the flow with its pinned procedures to a bundle file (default .sitelooper/procedures.json in cwd). Closes the session's browser. Then call build_start with the flow name.

Save a screenshot to disk and return its path (for evidence; you cannot see images).

Scroll an element into view.

Choose an option in a <select>, matching by visible label first, then by value.

List browser sessions and whether each is running.

Toggle network offline mode.

Set a session variable that later instructions reference as {{name}}; in a learn session it becomes a flow variable that build/check bind with vars. Not for secrets (use {{env:NAME}} markers).

Resize the viewport.

Accessibility snapshot of the current page with @ref handles for elements. Returns interactive/labelled elements only by default (what you need to pick something to act on); pass full:true for the complete tree including static text nodes. Call after navigation or DOM changes; refs from older snapshots go stale. For a specific value prefer read/read_all over a full tree.

Stop a browser session (closes its browser and daemon) without saving a flow. Use save_flow instead to keep a learn session's recording.

List open tabs, or switch the active tab by index.

Type text key-by-key into an element (triggers per-keystroke handlers, e.g. autocomplete). A {{env:NAME}} secret marker or a {{totp:NAME}} one-time-code marker in text is resolved at execution time — pass it through verbatim.

Set files on a file input.

Wait for a condition on a selector: visible, hidden, text_equals, text_contains, or count. Use this instead of sleeping or polling. Returns immediately if the condition already holds. Use count/text only for a value you expect to CHANGE — counting rendered rows is unreliable on virtualised lists (only visible rows exist in the DOM), so wait on a stable indicator instead.