Browser engine for AI agents with structured SOM output, V8 JavaScript, native MCP/AWP, and a supported CDP subset.
Plasmate provides 15 tools with mostly complete schemas and descriptions, but exhibits significant gaps in consistency, parameter documentation, and error handling guidance. Tool names follow verb_noun convention (fetch_page, click, type_text), which is good. However, descriptions vary widely in quality, some are detailed (fetch_page: 200+ chars with selector examples), others are minimal (extract_text: 60 chars, no guidance on when to use vs fetch_page). Parameter descriptions are present but often lack constraints, ranges, or format specifications. No tool includes error recovery guidance or actionable error messages. Session-based tools (open_page, click, type_text) lack clear documentation of session lifecycle and cleanup. Output schemas are not documented, LLMs cannot predict what fields to expect from responses. Duplicate tool sets (fetch_page vs plasmate_fetch, click vs plasmate_click) suggest integration-specific variants without clear differentiation. No tool declares permissions, security scopes, or audit requirements.
Discover ARD (Agent Resource Discovery) endpoints on a page.
Click an interactive element on the current page by its SOM element ID. Returns the updated page structure.
Close a persistent browser session.
Retrieve the crawl policy for a URL.
Run JavaScript code in a persistent browser session and return the result.
Extract plain text from a web page without semantic structure. Returns text content only.
Fetch a web page and return its semantic structure (navigation, content, forms, interactive elements). Returns a compact text representation with element IDs that can be referenced by other tools. Input: the URL to fetch, with optional budget, JavaScript, and SOM selector controls such as 'main' or 'interactive'.
Duplicate tool sets (fetch_page + plasmate_fetch, click + plasmate_click, type_text + plasmate_type) without clear differentiation or composition guidance. LLMs cannot determine which variant to use.
Session-based tools (open_page, navigate_to, click, type_text, evaluate, close_page) lack documentation of session lifecycle, timeout behavior, and cleanup requirements. No guidance on when sessions expire or how to handle orphaned sessions.
No output schemas documented. Tools return unspecified structures, LLMs cannot predict response fields, forcing trial-and-error chaining. fetch_page and inspect_page return 'semantic structure' but field names and types are undocumented.
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Inspect a page with visual rendering and return both SOM and screenshot.
Navigate to a URL in an existing persistent browser session. Maintains cookies and session state.
Open a URL in a persistent browser session. Returns the page's semantic structure with element IDs and a session ID for subsequent interactions.
Click an interactive element on the current page by its SOM element ID (e.g. 'e_a1b2c3d4e5f6'). Returns the updated page structure. You must call plasmate_navigate first to open a page.
Fetch a web page and return its semantic structure (navigation, content, forms, interactive elements). Returns a compact text representation with element IDs that can be referenced by other tools. Input: the URL to fetch, with optional budget, JavaScript, and SOM selector controls such as 'main' or 'interactive'.
Navigate to a URL in a persistent browser session. Returns the page's semantic structure with element IDs. Use this before click or type tools. Input: the URL to navigate to.
Type text into a form input or textarea on the current page. Requires the SOM element ID and the text to type. You must call plasmate_navigate first to open a page.
Type text into a form input or textarea on the current page by its SOM element ID.
Parameter descriptions lack constraints, ranges, and format specifications. 'budget' (fetch_page) has no min/max. 'selector' accepts free-form strings with examples ('main', 'interactive') but no enum or pattern. 'javascript' defaults to true with no explanation of performance impact.
No error handling guidance. Tools provide no recovery hints, error categorization, or actionable messages. E.g., if click fails because element_id is invalid, LLM receives no guidance on whether to retry, call fetch_page again, or abort.