No official connector
Google doesn't offer an MCP server or connector for coding agents like Claude Code or Codex. We found none from Google as of November 2026, so your agent can't reach your video credits on its own.
For Claude Code, Codex and any MCP agent · one-time payment
Your agent makes the video.
You stop copy-pasting.
Content by Agent connects your agent to the AI video credits you already have. Send it a video you love, it rebuilds it, then it cleans the watermark. One private repo. Pay once.
Requires Claude Code, Codex or another MCP client · One-time payment · Unofficial, not affiliated with Google. Read the limits first.
credits this session0
14MCP tools for your agent
0credits per image, at November 2026 prices
8models: 3 for images, 5 for video, all read live from your account
0subscriptions. You pay once.
Watch it
The whole thing in
ninety seconds.
With sound. Illustration of a Claude Code session; Claude Code is a product of Anthropic and this product is not affiliated.
The problem
You have the credits.
You can't use them from your agent.
Everything by hand
Copy a prompt, paste it, wait, download, rename, repeat for every shot. The credits are yours, but the time is spent on clicking.
A watermark you didn't choose
India
South Korea
Vietnam
Google's help page says a visible watermark is applied automatically to videos for users in these three countries unless they are on the Google One AI Ultra 20x plan. Source: Google's help page
| By hand | With Content by Agent | |
|---|---|---|
| Study a reference video | Watch it and describe it | Agent reads cuts, frames and contact sheets |
| Generate frames and clips | Copy and paste every prompt | Agent runs it through MCP tools |
| Know the cost | Find out after | Priced before every video, with a credit ceiling |
| Download and rename | Click, one file at a time | One tool call |
| Clean the visible watermark | A separate tool or a manual edit | Built in |
How it works
Install once.
Then just send a video.
- 01
Install
Accept the private repo invite, clone it, run
npm installand add it to your agent with one command. - 02
Sign in once
A dedicated Chrome window opens. You sign in to your Google account by hand. The tool never sees your password.
- 03
Send a video
Give your agent an mp4 and say what should change. It does the reading, the generating, the pricing and the cleaning.
# clone the private repo you were invited to
$ git clone <your repo link> contentbyagent
$ cd contentbyagent && npm install
# add it to your agent
$ claude mcp add contentbyagent -- \
node /path/to/contentbyagent/bin/contentbyagent.js
# check ffmpeg, python, chrome and sign-in
$ node bin/contentbyagent.js doctorYou type
@inspiration.mp4 remake this with a red dump truck
You get
- Your video's cuts, frames and contact sheets, read shot by shot
- A specific prompt for every shot, checked against a quality gate
- Frames generated and reviewed before any credit is spent on video
- Every clip priced first and refused above your limit
- A downloaded video with the visible watermark cleaned
What's inside
Four parts.
One private repo.
Video reader
video_read opens any video ffmpeg can read (mp4, mov, mkv, webm, avi, gif) and gives your agent its cuts, frames and contact sheets.
- Reads pictures, not speech
- Needs ffmpeg, not the browser
Studio tools
Twelve studio_* tools: open projects, generate images and videos, check the price, rename, download and upload.
- Spend guard on every video job
- Works in your own signed-in Chrome
Watermark engine
video_clean finds the visible sparkle in every frame, inpaints it and re-encodes the video with its audio in sync.
- Tuned for 9:16 (720×1280, 1080×1920)
- Position setting for other formats
The playbook
A skill that teaches your agent the workflow: prompt quality gate, frame and clip review rules, and what to do when a job is blocked.
- Built from full reel rebuilds (44 and 41 clips)
- Stops on bot checks, never loops
Models, lengths and prices
Every model your account lists.
Priced before every job.
The tools read the model list from your account, so a new model shows up without an update. Prices below were read from a Pro-plan account on 7 November 2026. Google changes models and prices, so your agent reads the live price before every video job.
Images
| Model | Credits | Aspect ratios | Tested |
|---|---|---|---|
| Nano Banana Pro | 0 | 16:9 · 4:3 · 1:1 · 3:4 · 9:16 | Image generated |
| Nano Banana 2.1 | 0 | 16:9 · 4:3 · 1:1 · 3:4 · 9:16 | Image generated |
| Nano Banana 2 Lite | 0 | 16:9 · 4:3 · 1:1 · 3:4 · 9:16 | Image generated |
1 to 4 images per job. Reference images by name are supported. Nano Banana 2.1 also has a thinking level in the studio; the tools use whatever is selected there.
Video
| Model | Length | Resolution | Credits | Tested |
|---|---|---|---|---|
| Omni 1.2 Flash newest | 3 to 15 s, in 1 s steps | 360p · 720p | 720p: 9 (3 s) · 12 (4 s) · 30 (10 s) · 60 (15 s) 360p: 3 (3 s) · 4 (4 s) · 10 (10 s) · 20 (15 s) | Priced and generated |
| Omni 1.1 Flash Leaving Dec 5 | 4 · 6 · 8 · 10 s | 360p · 720p | 720p: 7 · 10 · 12 · 15 360p: 4 · 5 · 6 · 7 | Priced and generated |
| Veo 3.1 Lite | 8 s | 720p | 10 | Priced and generated |
| Veo 3.1 Fast Leaving Dec 5 | 8 s | 720p | 20 | Priced and generated |
| Veo 3.1 Quality Leaving Dec 5 | 8 s | 720p | 100 | Priced and generated |
- Modes. Every video model offers Frames (a start and an end image) and Ingredients (reference images). The tools support both.
- Aspect. Video jobs run at 9:16 or 16:9.
- Upscale. Downloading at 1080p upscales the clip on demand. It showed 5 credits on the 15 s Omni 1.2 Flash clip we tested.
- Tested means a real job finished on our account, not only that the model was set up and priced. "Image generated" and "Priced and generated" are both real results.
- Spend guard. A job is refused when its price is above your limit (default 20 credits), so a 15 s Omni 1.2 clip needs you to raise the limit on purpose.
Real output
Same frame.
Drag to compare.
The corner of a raw 720p clip, enlarged 3×, before and after the watermark engine. A soft smudge can remain where the star was; it is not visible on moving footage at phone size.
It removes the visible sparkle only. Google's invisible SynthID mark stays in the file, and we recommend you use each platform's AI label when you post. More limits
BeforeAfter
Read this before you buy
What it doesn't do.
- It is unofficial. It automates Google Flow, in a Chrome window you signed in to yourself. It is not a Google product and does not use a Google API.
- Terms of service. Automating a website may go against its terms, and Google could limit an account it sees as misusing the service. You are responsible for your account.
- The invisible watermark stays. Only the visible sparkle is removed. SynthID remains in every file, so still use a platform's AI label when you post.
- Videos cost credits. Images cost 0 credits at current prices, but video clips spend credits from your own plan. This tool sells no credits.
- It depends on the web page. When Google changes the UI, a tool can fail until its selector is updated.
- Bot checks can stop a run. Google can show an "unusual activity" message. The tools report it and stop. They never try to get around it.
- One job at a time, Chrome open. Cloud agents that cannot reach your machine cannot use the studio tools.
- It reads pictures, not speech. Tell your agent about any voice-over or lyrics you want kept.
- Tested on Windows. The macOS and Linux paths exist in the code but are untested.
Pricing
Pay once.
No subscription.
Content by Agent
$49one time
Get Content by Agent
Preview price. Set PRICE and CHECKOUT_URL in app.js to go live.
- The MCP server: 14 tools for your agent
- The video reader,
video_read - The watermark engine,
video_clean - The playbook skill for Claude Code
- Delivered as a private GitHub repo, with a read-only invite to your account
- New commits land in the same repo you are invited to
You need a GitHub account, Node 18.17+, Google Chrome, Python 3.9+, ffmpeg (with ffprobe) and a Google account that can open Google Flow.
Questions
Short answers.
Is there an official connector from Google?
We found none as of November 2026. Community projects exist, and this is one of them, with a video reader and a watermark engine added.
Is the video generation free?
Not credit-free. Images cost 0 credits at November 2026 prices, but video clips spend credits from your own plan (a 4 s clip at 720p costs 12 on Omni 1.2 Flash and 7 on Omni 1.1 Flash). Google sets the prices and the daily allowance. This tool sells no credits and prices every video job before it runs.
Which countries get the automatic watermark?
Google's help page says India, South Korea and Vietnam, unless you are on the Google One AI Ultra 20x plan. Other users see a visible-watermark setting in their profile menu. Check Google's page for the current rule, because it can change.
Does it remove every watermark?
No. It removes the visible sparkle only. Google's invisible SynthID mark stays in the file, and we recommend using each platform's AI label when you post.
Does it work with Claude?
It runs as a stdio MCP server, so it works in Claude Code, Codex and any client that can run one. Chrome must stay open on the same machine.
Does it copy the video I send?
It reads the structure of the reference, such as cuts, shot types and pacing, and generates new footage with the changes you ask for. You are responsible for having the right to use any reference you send.
Could my Google account be limited?
It could. Automating a website may go against Google's terms, and Google can limit an account it considers to be misusing the service. Read Google's terms before you use it on an account you care about.
Which models does it work with?
Every image and video model your account lists. The tools read the list from the page, so new models appear without an update. See the Models section for lengths, resolutions and credit prices. We tested all three image models and all five video models on our own account.
Which formats does the watermark engine support?
It is tuned for 9:16 videos at 720×1280 and 1080×1920. For other formats you set the star's position once.
Which systems does it run on?
It was tested on Windows. The macOS and Linux Chrome paths are in the code but have not been tested yet.
How do I get the repo?
After checkout you get a read-only invite to the private GitHub repo, sent to your GitHub account.