Monorepo for bridging local MCP clients to browser WebMCP through Playwright, adapters, and native example apps
WebMCP Bridge exposes 28 tools across diagram, node, edge, selection, layout, canvas, view, auth, and search domains. While tool names follow verb_noun conventions (diagram.get, nodes.upsert, edges.remove), critical gaps undermine quality: (1) Many tools lack parameter descriptions entirely (e.g., edges.upsert items array has no schema details; layout.apply algorithm param undescribed). (2) Output schemas are not documented, LLMs cannot predict return structures. (3) Error handling is absent, no recovery guidance or categorization. (4) Several tools have minimal descriptions (<50 chars: 'Report whether this template adapter is ready', 'Echo an input value for connectivity checks'). (5) Destructive operations (nodes.remove, edges.remove, diagram.reset) lack confirmation patterns. Per-tool analysis: diagram.get, nodes.list, edges.list score ~55 (clear names, basic schemas, but no output docs); nodes.upsert, edges.upsert score ~40 (complex nested objects with undescribed fields); auth.get, echo.execute score ~25 (trivial descriptions, minimal schemas); gemini.chat, search.web score ~60 (better descriptions, but missing error guidance and output structure). Average across 28 tools: 42.
Report whether this template adapter is ready
Detect whether the current Google session still needs interactive sign-in
Patch visual styles for the canvas background.
Export the current diagram to a file.
Get the current structured diagram snapshot derived from the authoritative Excalidraw scene.
Replace the current diagram with the built-in bridge architecture demo scene.
Reset the diagram to an empty state.
Output schemas not documented. LLMs cannot predict return structures for any tool. Diagram.get, nodes.list, edges.list, and others lack documented response schemas, forcing LLMs to guess at field names and types.
Complex nested parameters undescribed. edges.upsert accepts an 'edges' array with items of type 'object' but provides no schema for edge properties (source, target, style, etc.). nodes.upsert similarly lacks detailed field documentation for the nested 'nodes' array.
Destructive operations lack confirmation or dry-run patterns. nodes.remove, edges.remove, and diagram.reset can permanently delete data without any confirmation step, undo capability, or recovery guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Set the current diagram title used by the board UI and export file names.
Echo an input value for connectivity checks
List all structured diagram edges.
Remove one or more edges by id.
Patch visual styles for one or more edges without changing their structure.
Create or update one or more diagram edges. New edges require source and target node ids. Existing edges can be patched by id.
Send a Gemini prompt for text or image generation and optionally download generated images
Download generated Gemini images from the current or target conversation
Apply an automatic layout algorithm to the diagram.
List all structured diagram nodes with positions and kinds.
Remove one or more nodes by id and delete their dangling edges.
Resize one or more nodes without changing their positions.
Patch visual styles for one or more nodes without changing their structure.
Create or update one or more architecture nodes. New nodes require label and kind. Existing nodes can be patched by id.
Return the current browser page URL and title for debugging
Inspect interactive elements on the current page for adapter development
Navigate the current browser page to a Google-owned URL
Run a Google web search and return visible result cards
Get the current selection state including selected node and edge ids.
Clear the current selection.
Fit the view to the diagram content or selection.
Trivial descriptions on utility tools. auth.get ('Report whether this template adapter is ready'), echo.execute ('Echo an input value for connectivity checks'), and layout.apply ('Apply an automatic layout algorithm') lack context for when/why LLMs should call them. Descriptions under 50 chars cannot guide tool selection.
No error handling or recovery guidance. Tools provide no error categorization (retryable vs fatal), no actionable error messages, and no hints for recovery. An LLM receiving a failure has no guidance on next steps.