Google Flow Generate

Google Flow Generate

Runbook to generate images/video on Google Flow through the google-flow MCP server. Set FLOW_MCP_DIR below to wherever you cloned this repo.

STEP 0 — Ensure the dedicated Chrome is up (always run first)

powershell -NoProfile -ExecutionPolicy Bypass -File "/scripts/ensure-flow-chrome.ps1"

• READY / LAUNCHED → proceed.
• LAUNCHED: if labs.google shows the marketing landing instead of the app, the user

must click "Sign in to Flow" once in that Chrome window, then continue.

• FAILED → report; do not proceed.

If the mcp__google-flow__* tools are absent, the MCP server needs a client restart (it loads into RAM at startup).

STEP 1 — Connect

Call flow_connect:

• status: connected → go on.
• status: oauth_required → user signs in to Flow once, then retry.
• WRONG_GOOGLE_ACCOUNT → stop; setup mismatch.
• flow_account_check → method:"assumed" means unverified, not a green light.

STEP 1.5 — Build a faithful prompt (always, before generating)

Flow's agent and the models render far more faithfully from a rich, structured prompt than from a bare phrase. Turn the user's request into a complete prompt without betraying intent — enrich, never invent:

• Keep every element the user named; never drop or swap any.
• Add only supporting detail (subject + action + setting + lighting + composition +

style/medium + mood + quality cues; for video also camera movement + pacing). Do NOT add subjects/objects/text the user didn't ask for. Preserve explicit constraints ("red mug, no logo") verbatim.

• Match ratio to the use (16:9 scene, 9:16 story, 1:1 icon).

The server also wraps the prompt so the agent stays bound to it verbatim.

STEP 2 — Images (free)

flow_generate_image with auto_confirm:true. Models: Nano Banana 2 (default), Nano Banana Pro (best fidelity — prefer for complex prompts), Imagen 4. Params: prompt, ratio, project_name, campaign. Result files[] = saved paths. (The credits_consumed flag is hardcoded; images are free against the Flow pool.)

Verify & refine (images — free, so always do it)

After generating, open the image with vision and compare it to the request element by element (subjects present? constraints respected? right composition? nothing unwanted added?). If it matches → deliver. If not → refine the prompt to emphasize the failed element and regenerate (up to 3×; images cost nothing). This catches real misses — e.g. the download picking a UI asset instead of the generated image.

STEP 3 — Video (⚠️ consumes credits)

flow_generate_video with auto_confirm:true submits, approves the credit dialog, waits ~1 min and downloads the mp4.

• Credits: Veo 3.1 Lite ~10, Fast ~20, Quality ~100; Omni Flash ~15-30 of ~1000/month.
• Valid model/duration combo required (Veo 3.1 Lite = 8s-only on Pro; Omni Flash

4-10s) — else the agent asks for clarification and nothing generates.

• Confirm with the user before spending credits if not already authorized.
• auto_confirm:false stages without spending.

Fallback map (error → action)

| Error | Action | |---|---| | oauth_required, MANUAL_VERIFICATION_REQUIRED | STOP, user signs in / solves captcha | | WRONG_GOOGLE_ACCOUNT | STOP, setup problem | | BROWSER_NOT_CONNECTED, NOT_LOGGED_IN, FLOW_PAGE_NOT_FOUND | rerun STEP 0, then fall back | | GENERATION_BUTTON_DISABLED, GENERATION_TIMEOUT, UNKNOWN_UI_CHANGE | fall back + flow_discover_ui | | GOOGLE_LIMIT_REACHED | credits exhausted → fall back | | DOWNLOAD_FAILED | don't regenerate (credits spent) — retry download |

Image fallback: Flow → Gemini API (fallback/gemini_image_fallback.py, needs GEMINI_API_KEY) → any per-credit generator you use. Video fallback: Flow → your per-credit generator directly.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: GabrielGargiuloDev
• Source: GabrielGargiuloDev/google-flow-mcp
• License: MIT

Install and usage instructions live in the source repository linked above.