A Model Context Protocol server for browser automation, webpage interaction, and workflow recording/replay using Chrome DevTools and extension integration
Webpage MCP has 49 tools with significant structural problems. While naming is consistently verb-noun (get_windows_and_tabs, recording_start, chrome_navigate), many descriptions are terse or missing critical details. Input schemas are visible for most tools but many parameters lack descriptions. The server mixes high-level workflow tools (recording_start, flow_analyze) with low-level browser automation (chrome_click_element, chrome_javascript), creating composition hazards. Critical gaps: no pagination guidance for list operations, inconsistent parameter naming (tabId vs flowId), missing error recovery documentation, and no output schema definitions visible in the tool definitions. The error handling pattern is absent, tools fail silently rather than guiding LLM recovery. Security concerns: chrome_inject_script and chrome_javascript expose arbitrary code execution without sandboxing or permission gates.
29 of 49 tools (59%) have NO visible description in the provided source. Tools like search_tabs_content, chrome_navigate, chrome_screenshot lack any description text, making tool selection impossible for LLMs.
Add descriptions to all 29 missing tools immediately. Each description should be 50-200 characters and answer: What does it do? When should the LLM call it? What prerequisites exist? Example: 'chrome_click_element: Click a DOM element by CSS selector or XPath. Use after chrome_get_interactive_elements to locate the element. Returns success/failure and any post-click state changes.'
Document input schemas for all 29 tools lacking visible parameters. Even if parameters are minimal (empty object {}), explicitly declare the schema and any constraints. Example: chrome_navigate should declare {url: {type: string, description: 'Target URL to navigate to'}, required: ['url']}'
Add output schema documentation to every tool. Include at minimum: field names, types, and what values are safe for downstream tool chaining. Example: get_windows_and_tabs should document {windows: [{id: number, tabs: [{id: number, title: string, url: string}]}]} so LLMs know tabId is available for chrome_switch_tab.
Standardize parameter names across related tools: use 'tab_id' everywhere (not 'tabId', not 'tabId'). Use 'flow_id' everywhere. Follow snake_case consistently.
Add explicit WRITE/DESTRUCTIVE declarations to tool descriptions for all state-modifying operations. Example: 'chrome_close_tabs: DESTRUCTIVE. Close one or more tabs by ID. This action cannot be undone. Requires user confirmation in multi-tab scenarios.'
Implement pagination and result limits for list tools: add limit (default 20, max 100) and offset/cursor parameters. Document that results are capped. Example: 'record_replay_list_published: List all published workflows (max 100 per request). Returns {workflows: [...], total_count: N, has_more: boolean, next_cursor: '...'}'
Register and manage persistent user scripts in the browser.
flow_analyzeread onlysource verified58/100
Analyze a recorded flow and return optimization hints plus a minimized public flow view (summary structure, sanitized variables, and safe metadata only).
flow_updatewritesource verified60/100
Update a flow using partial fields such as nodes, edges, variables or metadata.
Cancel a queued, running, or paused record/replay run by runId. Queued runs are canceled immediately; active runs receive a cancel request and can optionally be waited to terminal state.
recording_startwritesource verified65/100
Start recording user interactions on an HTTP(S) browser tab. If tabId is omitted, uses the active tab.
recording_statusread onlysource verified65/100
Get the current recording status and session snapshot.
recording_stopwritesource verified63/100
Stop the active recording session and persist the recorded flow. Optionally rename the flow.
search_tabs_contentread onlysource verified25/100
Search across all tab content by keyword or phrase. Returns matching tabs with context snippets.
workflow_approval_storewritesource verified25/100
Store and manage workflow approval tokens for sensitive operations.
workflow_debug_viewread onlysource verified27/100
Get debug information about a workflow run including execution state, errors, and artifacts.
workflow_describeread onlysource verified27/100
Get detailed information about a workflow including its structure, variables, and execution requirements.
workflow_migratewritesource verified27/100
Apply runtime migrations to update a workflow to a new runtime version.
workflow_publishwritesource verified28/100
Publish a workflow to make it available for general use.
29 tools have NO visible input schemas in the provided source code. Even though the tool definitions exist in packages/shared/src/tools.ts, their input parameter schemas are not shown, making schema validation impossible to assess.
Tools like chrome_inject_script, chrome_javascript, and chrome_send_command_to_inject_script expose arbitrary code execution without visible permission gates, rate limits, or sandboxing. No security pattern documentation visible. These are security-critical tools that require explicit authentication and audit trails.
No output schemas documented for any tool. LLMs cannot plan chaining (what fields to extract for downstream calls). For example, get_windows_and_tabs should return tab IDs usable by chrome_switch_tab, but no documentation shows the response structure.
Tools that modify state (recording_start, chrome_click_element, workflow_publish) have descriptions that do not explicitly declare they are WRITE or DESTRUCTIVE operations. LLMs cannot determine if a call is retryable or has irreversible consequences.
No pagination, limit, or result-capping documented for list operations. Tools like record_replay_list_published and search_tabs_content likely return unbounded results, risking context window exhaustion and LLM reasoning degradation.
No error recovery guidance visible. Tools fail without actionable error messages that tell the LLM what to do next. E.g., if workflow_repair fails, the LLM has no guidance on alternative repair strategies or manual escalation.
Add error recovery guidance to all tools. When a call fails, return {error: {code: string, message: string, actionable_next_step: string, retryable: boolean}}. Example: 'workflow_repair failed because syntax error at line 12. Try: (1) inspect with workflow_debug_view, (2) manually fix in UI, (3) retry repair.'
Sandbox chrome_inject_script and chrome_javascript with explicit permission checks and rate limits. Document required scopes: 'Requires scope: write:tabs:javascript. Maximum 10 executions per minute. All code is logged for audit.'
Split tools with multiple concerns. For example, if flow_analyze both analyzes AND returns a sanitized view, split into flow_analyze (optimization hints) and flow_export_sanitized (public view).
Add parameter descriptions to all tools with visible schemas. Every parameter needs a 20+ character description. Example: flow_update's 'applyParameterSuggestions' should explain: 'If true, automatically applies recorded parameter suggestions to nodes and creates missing variables. Safe to use for auto-refinement workflows.'