flow-mcp
flow-mcp lets your AI assistant drive Google Flow through your own signed-in Chrome session to create and manage images, videos, characters, and projects without an API key.
Create or edit Nano Banana images, including parallel batches, for free.
Generate Veo 3.1 and Omni videos from text, images, ingredients, or source videos, with credit-budget approval before spending.
Upscale/download media at 1K, 2K, or 4K; collect finished renders and extract video frames.
Manage Flow projects, assets, duplicates, and two-step media trashing.
Create and reuse consistent Flow characters with reference images, notes, and voices across generations.
Check credits, sign-in/health status, Flow capabilities, project settings, and agent transcripts.
Chat with Flow's creative agent, approve or reject credit spends, and wait for generated media to finish.
Drives Google Flow (flow.google.com) in a signed-in Chrome session to generate and edit images, create Veo/Omni videos, upscale media, and interact with Flow's creative agent — using the user's own Flow plan and credits, with credit checks and spend approval before paid generations.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@flow-mcp generate a photorealistic image of a lighthouse at dusk with a calm sea"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
flow-mcp 🎬 — Google Flow MCP server for Claude, Cursor and VS Code
flow-mcp is an open-source MCP server that lets your AI assistant make images and videos with Google Flow. Ask Claude Code, Cursor, VS Code or Claude Desktop in plain words and it generates and edits Nano Banana images, makes Veo 3.1 and Omni videos, upscales to 2K, keeps a character consistent across shots, and works with Flow's creative agent.
Your account, your credits, no catch. flow-mcp drives Flow in a Chrome window signed in to your own Google account, on your own computer. No API key, no server in between, no extra subscription. Images are free, and videos only run within the credit budget you set.
Getting started · Usage · Gallery · Capabilities · Troubleshooting · Security · npm
Install
You need Node.js 22+ and Google Chrome. A free Google account works: it gets 50 Flow credits a day.
npx -y @parthmax/flow-mcp setupA Chrome window opens on Flow; sign in to Google once. setup then offers to connect Claude Code and
install the flow-director skill, and prints the config for Cursor, VS Code and Claude Desktop. Restart
your AI client afterwards.
Or let your assistant do it. Paste this to Claude Code:
Set up the Google Flow MCP server for me: run
npx -y @parthmax/flow-mcp setup --connect. A Chrome window will open; tell me to sign in to Google there. When it is done, tell me to restart you.
Client configs and running from source are in Getting started.
Related MCP server: google-flow-mcp
Quick start
Ask your assistant:
make an image of a lighthouse at dusk
edit my photo C:\photos\me.jpg into a winter scene
make an 8 second video of a paper boat, max 10 creditsA Chrome window shows Flow doing the job; leave it open. Files are saved to ~/.flow-mcp/output/<project_id>/. If
something fails, npx -y @parthmax/flow-mcp status names the problem and the fix.
Made with flow-mcp
Everything below was made by an AI assistant through flow-mcp, from one mascot reference. The images were free and the video cost 10 credits. Same character in every shot, thanks to a character sheet and Flow's references.
How it fits together
The MCP server (
flow-mcp serve) runs on your computer and gives your assistant 20 Flow tools: images, videos, upscaling, characters, projects, credits and Flow's agent chat.A Chrome window with its own profile stays signed in to your Google account. flow-mcp drives Flow's own page there, so Flow handles sign-in and verification as it does for you.
Projects keep each piece of work together: Flow's agent remembers it, and local files upload once and are reused.
The flow-director skill plans shots, previews everything as free images, then spends credits once the plan is right.
See How it works for the details.
Credits and safety
Images, edits and 2K upscales are free. For videos, Flow's agent quotes a price first, and flow-mcp
approves only within your max_credits (default 10). It never picks "Always approve" and never retries a
paid generation. An 8 s Veo 3.1 Lite video costs 10 credits on a free account (October 2026).
Alpha, unofficial. Tested on Windows with an English-language Flow account. Automating the Flow website may be against Google's terms, and the local control port can act as your Google account. Read Security and risks first. Not affiliated with Google.
Documentation
Goal | Start here |
Install and connect a client | |
Make images and videos, manage credits | |
Get better results from Veo and Nano Banana | |
Change folders, ports, or Google accounts | |
Fix a problem | |
Understand the risks | |
Learn how it is built | |
Give an AI the project summary |
Development
git clone https://github.com/parthmax2/flow-mcp.git
cd flow-mcp
npm install
npm run build
npm testCoding agents working on this repo: read AGENTS.md.
Community
Use GitHub issues for bugs and feature requests. If Flow's
website changed and something broke, include the status output. Fixes, macOS and Linux testing, and new
tools are welcome as pull requests.
Built and maintained by Saksham Pathak (Parthmax): parthmax.in · LinkedIn · Hugging Face
License
MIT. See LICENSE. Google and Google Flow are trademarks of Google LLC.
Available Tools
18 toolsflow_approveApprove or reject the Flow agent's spend request (may spend credits)A
Answers the agent's pending 'costing N credits?' request. decision 'approve' SPENDS the quoted credits, but only when N <= max_credits; otherwise it rejects and nothing is spent. Never chooses 'Always approve'. Returns the agent's reply and the media it started (videos then render for 1-5 min: fetch them with flow_wait using since/media).
Inputs: decision (required), project_id (required), max_credits (required), include_thinking.
flow_assetsList the media in a Flow projectARead-only
Lists a project's images and videos, newest first: id, title, image/video, uploaded/generated, model, prompt, size, duration and what each generated item was made from. Attach any of them to a request with media_ids instead of uploading files again. duplicateUploads lists repeated uploads of the same file, with the copy to keep.
Inputs: project_id (required), include_trashed.
flow_capabilitiesWhat Google Flow and this server can doARead-only
Lists Flow's features, which tool covers each (or that only the Flow website can), costs and plan limits. Free.
Inputs: no named parameters.
flow_charactersList, create or change Flow charactersA
Flow's Characters keep a person (or creature) the same across generations: a reference image, a name, notes the Flow agent reads, and a voice. action 'list' (default) returns the project's characters. 'create' makes one from an image already in the project (image_id, e.g. an approved character sheet or keyframe) and sets name, info and voice. 'update' changes name, info or voice of an existing one (character: name or id). 'voices' lists Flow's voices (needs one character in the project). Use characters in generate_video, generate_image or flow_chat with characters: [name]. Free.
Inputs: project_id (required), info, name, voice, +3 more.
flow_chatTalk to the Google Flow agentA
Sends one message to the Flow agent and returns as soon as it has finished replying, with its exact words (messages, reply) so you can decide the next step: answer its questions, refine the request, or handle a spend approval. Sending never spends credits by itself. If the reply has pendingApproval (a quoted credit cost), the agent is waiting: tell the user the cost and use flow_approve. Images are usually made right away (no approval); their ids are in media, download them with flow_wait.
Inputs: message (required), media_ids, characters, project_id, +6 more.
flow_creditsGoogle Flow credit balanceARead-only
Reads the account's remaining Flow credits. Free accounts get 50 a day; an 8 s Veo 3.1 Lite video costs 10.
Inputs: no named parameters.
flow_frameTake a still frame from a videoARead-only
Saves one frame of a video as a PNG: the first, the last (default) or one at a given second. The video is a local file or a project media id (with project_id; downloaded first if needed). Main use: pass the last frame of a shot as start_frame of the next generate_video, so characters and setting carry over between clips. The PNG is uploaded when used, so Flow asks for the rights confirmation then. Free; runs on this machine.
Inputs: video (required), at, project_id.
flow_projectsList Google Flow projectsARead-only
Lists the account's Flow projects, most recently updated first, with the purpose and upload count this tool recorded for projects it created. Use it to decide whether to continue a project or start a new one.
Inputs: no named parameters.
flow_settingsRead or change a Flow project's agent settingsA
Without changes: returns the project's agent settings (default aspect, count and model for images and videos, whether the agent asks before spending credits), the models this account offers, and its standing instructions. With changes: sets them and returns the result. Use it to avoid surprises, e.g. image count 1 instead of the default 2, or to pick a video model. Approval can only be set to 'always' (the credit budget depends on it). instructions replaces all standing instructions (an empty list removes them). Free.
Inputs: project_id (required), image, video, approval, +1 more.
flow_sign_inSign in to Google FlowA
Opens the Flow Chrome window (starting it if needed), brings it to the front, opens Flow's Google sign-in and waits until the user has signed in there, then returns the account. Use it when flow_status says 'signed_out' or on first use: first tell the user a Chrome window will open and they should sign in with their Google account. Returns at once if someone is already signed in. Free.
Inputs: wait_seconds.
flow_statusGoogle Flow health checkARead-only
Checks that everything works and names the problem if not: state 'ready', or 'browser_down', 'signed_out', 'challenge' (Google wants a verification), 'blocked', or 'ui_changed' (Flow's page changed; lists which controls are missing). Includes the signed-in account, credits and a fix to tell the user. Run it first, and whenever a tool fails unexpectedly. Free; takes about 30 s.
Inputs: no named parameters.
flow_transcriptRead a Flow agent conversationBRead-only
Returns the whole agent chat of a project (both sides, approvals and whether they are still pending). Free.
Inputs: project_id (required), include_thinking.
flow_trashMove Flow media to the trash (two steps)A
Moves project media to Flow's trash, where it can still be restored in Flow. Always two calls: first without confirmation, which only previews the items (title, type, what was made from them) and returns a confirmation code; show that list to the user and get their OK; then call again with the same media_ids and that code. The code only works for exactly the previewed set. To clean up repeated uploads use duplicateUploads from flow_assets.
Inputs: media_ids (required), project_id (required), confirmation.
flow_upscaleUpscale or download Flow media at a larger sizeA
Downloads a project's media at the chosen size, upscaling it in Flow: images 2K (free on every account) or 4K (paid plans). Without size, lists the sizes Flow offers for that item on this account, including locked ones. Offer the free 2K upscale to the user after new images rather than doing it unasked.
Inputs: media_id (required), project_id (required), size.
flow_waitCollect Flow media when it is readyA
Checks on media the agent is making (e.g. a video from generate_video or flow_approve) for up to timeout_seconds, downloads whatever is finished and returns anything the agent said meanwhile. status 'completed' means all of it is downloaded; 'pending' means it is still rendering: tell the user and call again (videos usually take 1-5 minutes); 'failed' means Flow reported the generation as failed (failures has its message): tell the user, suggest checking flow_credits for the refund, and do not keep waiting. Pass media_ids, or since from the earlier result to take all new media. Free.
Inputs: project_id (required), since, media_ids, timeout_seconds.
generate_imageCreate or edit images with Google FlowA
Asks the Flow agent for images and saves them locally. To edit, attach the image (attachments for local files, media_ids for project media) and describe the change, or in the same project just name the image the agent made. The agent may rewrite the prompt and make several images (default 2 with Nano Banana); say 'one image' to steer it. The result includes what the agent said. Images use no Flow credits. Takes 30-120 s.
Inputs: prompt (required), media_ids, characters, project_id, +5 more.
generate_imagesCreate several images at once with Google FlowA
Runs several separate image requests in parallel (up to 3 Flow tabs at a time, each in its own agent session), about 3x faster than one by one. Use it for a set of different images, e.g. one per scene or one per reference frame. Files in attachments are uploaded once for the whole batch. Each job reports its own files or error. Free.
Inputs: jobs (required), parallel, project_id, new_session, +3 more.
generate_videoStart a video with Google Flow (spends credits)A
Shortcut for flow_chat + flow_approve: asks the agent for a video and approves its quote only when it is at or below max_credits (default 10 = one 8 s Veo 3.1 Lite video on a free account), otherwise rejects. SPENDS FLOW CREDITS: tell the user the expected cost and check flow_credits first; never call it again to retry. Returns as soon as the render has started: 'started' (collect the mp4 with flow_wait, passing media ids or since), 'needs_answer' (the agent asked something: show the user and reply with flow_chat), 'over_budget' (nothing spent) or 'failed' (Flow failed the render right away; failures has its message). Every result includes the agent's messages. A source_video makes it an Omni video-to-video edit (20 credits at 720p, same length as the clip). Inputs given a role (start_frame, end_frame, ingredients, source_video: each a local file or a project media id) are numbered and explained to the agent; prefer them to plain attachments. Veo takes either frames (image-to-video) or ingredients (reference-to-video), not both: given both, the agent asks which to use (needs_answer, nothing spent).
Inputs: prompt (required), model, end_frame, media_ids, +12 more.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 18 tools
Most tools map to a distinct resource or step (status, sign-in, credits, projects, assets, trash, characters, settings, wait, approve), and descriptions explicitly disambiguate tricky pairs like generate_video vs. flow_chat/flow_approve and generate_image vs. generate_images. The only real friction is the intentional overlap between the video shortcut and the chat+approve workflow, plus the single-vs-batch image pair, which an agent must read carefully to choose correctly.
Almost all tools carry the flow_ namespace prefix, which is predictable, but the pattern is not verb_noun: several are bare nouns (flow_status, flow_credits, flow_projects, flow_settings) while the core generation tools break the namespace entirely (generate_video, generate_image, generate_images). Readable, but the convention is mixed rather than uniform.
18 tools is on the heavy side but each covers a genuinely separate part of driving a browser-based generation service (auth, credits, assets, generation, approval, polling, cleanup). Nothing looks redundant enough to cut, though a few convenience wrappers (flow_frame, flow_capabilities) push it slightly past a tight scope.
The lifecycle is well covered: health check, sign-in, credits, project listing, image/video generation, batch generation, upscaling, asset listing, trash with confirmation, characters, settings, chat/approve/poll and frame extraction. Gaps are minor but real: no project create/rename/delete, no character deletion action, and no direct asset download except through upscale or wait.
Maintenance
Related MCP Connectors
Let AI agents use your real, signed-in Chrome: tabs, clicks, forms, screenshots and video.
Connect AI agents to your own Chrome with scoped access, approvals, and verified task results.
Google Flow's agent: chat that plans and makes images and videos, approvals, your own media
Give AI agents a real cloud Chrome: open sessions, navigate, read pages, click, type and screenshot.
Related MCP Servers
- AlicenseAqualityDmaintenanceEnables AI agents to drive Google Flow through a real Chrome profile to generate images, videos, characters, and scenes without sharing credentials.1910 npm2MIT
- AlicenseAqualityBmaintenanceEnables AI agents to generate Google Flow videos and images and automate Scene Builder clip extensions through a user's own Chrome session.11264 npm10MIT
- AlicenseNot gradedqualityCmaintenanceEnables generating images and videos on Google Flow through browser automation using your Google AI Pro subscription, with image generation free and video consuming Flow credits.117 npm1MIT
- AlicenseNot gradedqualityCmaintenanceEnables generating images and videos in Google Flow through an automated browser session using your Google AI Pro subscription, without API keys or third-party credit fees.117 npm2MIT