Browser automation framework exposing DDM (DOM Density Map) and page intelligence tools as MCP tools for Claude Code and Claude API. Enables self-drive browser automation with CDP transport, DOM analysis, and Bayesian page intelligence.
33 tools with generally clear naming and descriptions, but significant gaps in schema completeness and parameter documentation. Most tools have action-verb names (ddm, navigate, click, screenshot) and 1-2 sentence descriptions. However, many parameters lack type constraints, enums, or validation guidance. Output schemas are not documented. Error handling is minimal. Tool composition is reasonable (single responsibility mostly observed), but some tools like cdp_command expose raw API surface without abstraction.
Execute a raw Chrome DevTools Protocol command.
List all frames in the current page, including iframes. Returns frame IDs and URLs for use with ddm_frame and js_eval_frame.
Click at coordinates or on an element. Coordinates are pixel positions on the page.
Close a browser tab.
Create a new browser tab.
DOM Density Map — structural page layout + interactive elements. Returns ~500 tokens of page understanding. Use this FIRST on every page to orient yourself before taking any action.
Output schemas not documented. Tools like ddm, intel_probe, screenshot, and stream_screencast lack documented return types. LLMs cannot plan downstream tool calls or extract specific fields without knowing what the response contains.
Parameter validation constraints missing. Many parameters (e.g., 'flags' in ddm, 'strategy' in intel_extract, 'direction' in scroll) accept free-form strings without enums or format constraints. LLMs will hallucinate invalid values.
cdp_command exposes raw Chrome DevTools Protocol without abstraction. Requires LLM to know CDP method names and parameter structures. Violates single-responsibility principle and increases error surface.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Run DOM Density Map inside a specific iframe. Use after cdp_list_frames to see iframe layout and interactive elements.
Get cookies from the current tab, optionally filtered by URLs.
Extract structured data using auto-selected or forced strategy. Strategies: innerText, host_attrs, js_global, react_fiber, data_testid, heading_hier, img_alt, shadow_pierce.
Find paths to a value in a data store using fuzzy matching. Use after intel_shape to locate specific data.
Page intelligence probe — DOM fingerprint + Bayesian strategy ranking. Returns ~100 tokens. Identifies the page framework (Nuxt/Next/React), data stores, shadow DOM structure, and ranks 8 extraction strategies.
Infer the shape of a data store (keys, types, nesting). Use after intel_stores to understand store structure before querying.
Enumerate data stores (Vuex, Pinia, Redux, Zustand, MobX, Jotai, Recoil) and their current state. Use after intel_probe ranks js_global high.
Execute JavaScript in the page context and return the result. Use for querying or manipulating the DOM.
Execute JavaScript inside a specific iframe. Use after cdp_list_frames to interact with iframe content.
Press a keyboard key with optional modifiers.
List all connected agents (browser profiles) for the authenticated user.
Navigate to a URL in the current tab.
Clean up a provisioned Chrome instance.
Launch a provisioned Chrome instance with a specific profile path.
Get the status of provisioned Chrome instances.
Scan page DOM for specific data terms using Rhythm. Use after rhythm_train to find data on the page.
Execute a multi-step plan on a page using Rhythm at event speed (no LLM between steps).
Query what Rhythm has already learned about a site without training again.
Train Rhythm on a page to learn element timing and event patterns. Run once per site before using rhythm_catch or rhythm_execute.
Capture a screenshot of the current page.
Scroll the page in a direction by an amount.
Set cookies in the current tab.
Set a file input element to a file path.
Stream screencast frames from the page at specified intervals.
Submit the focused form by pressing Enter.
Type text into the focused element.
Wait for the page to be ready using a specified strategy.
Error handling guidance absent. No tool descriptions explain what to do on failure, which errors are retryable, or how to recover. LLMs will not know whether to retry, ask the user, or abandon the task.
Numeric parameter ranges not specified. 'amount' in scroll, 'quality' in screenshot/stream_screencast, 'max_frames' in stream_screencast lack min/max bounds. LLMs may pass absurd values (e.g., amount=999999, quality=1000).