Loading ToolBench
MCP server for Google Flow (flow.google.com) via browser automation
Google Flow MCP has 19 tools with significant definition gaps. Most tools lack proper parameter descriptions (many params have no description field at all). Descriptions are present but often vague or overly technical (e.g., 'Prepare a Veo video (prepare-only, no spend)' lacks actionable context). Schemas are partially defined but incomplete, many parameters lack type constraints and descriptions. Error handling is minimal; no recovery guidance. Tool naming is reasonable (verb_noun pattern mostly followed) but some tools conflate multiple concerns (e.g., flow_generate_image handles both preparation and execution). No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels. Output schemas are not documented. Critical issue: parameters like 'project_name', 'campaign', 'reference_images' appear across many tools but lack consistent descriptions or validation guidance.
Verify the logged-in Google account matches the configured expected email (Default).
Launch Chrome with the configured Google profile, connect CDP, navigate to Google Flow, and verify account.
Create a new character in the current project.
Create a new scene in the current project.
Close the browser and clean up the MCP connection to Google Flow.
Navigate to a Google Flow page and discover all interactive elements (buttons, inputs, links, headings). Updates the internal selectors map for robust automation.
Download the latest generated image or video from the current project.
Missing parameter descriptions across all tools. Parameters like 'project_name', 'campaign', 'reference_images', 'params' lack descriptions or have trivial ones. LLMs cannot infer parameter meaning from names alone.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. flow_generate_image and flow_generate_video are marked DESTRUCTIVE (consume credits) but lack destructiveHint annotation to warn agents.
Output schemas not documented. Tools return complex objects (elements, screenshot, status) but LLMs have no schema to understand structure. E.g., flow_discover_ui returns 'elements' but type/structure undefined.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | <=2025-11-25 | v2 |
⚠️ THESE IMAGES CONSUME CREDITS. By default (auto_confirm=false): fills the prompt, selects model/ratio, takes a screenshot and returns "ready_for_confirmation". Does NOT click Generate. When auto_confirm=true: first verifies the UI is in IMAGE mode (not Video), that the model is an image model, takes a verification screenshot, THEN clicks Generate and waits for the images. NANO/BANANA image models only.
Prepare a Veo video (prepare-only, no spend). Fast tests, Quality finals; ingredients need 8s.
Import a character from a reference image or URL.
List available project images, characters, and scenes for use in prompts via @name references.
List all available tools in the tools gallery.
Navigate the already-connected browser to a Google Flow URL without reconnecting. Only https://flow.google.com/... URLs are allowed (other hosts rejected).
Open the characters section of a project.
Open a specific tool by name from the tools gallery.
Open the tools gallery in Google Flow.
Check current connection status: browser connected, Flow page loaded, account verified, job queue state.
Use a specific Flow tool with parameters.
Use the Grid Architect tool for layout design.
Vague descriptions that lack actionable context. E.g., 'Prepare a Veo video (prepare-only, no spend)' does not explain WHEN to call this vs flow_generate_image, or what 'prepare-only' means to an LLM.
No error handling guidance. Tools throw FlowError but descriptions do not explain recovery steps. E.g., if flow_open fails due to invalid URL, LLM has no guidance on what to try next.