no-api-media

no-api-media-mcp is an open source MCP server (Apache-2.0, Node 22+) that lets Claude Code, Cursor, Codex and other MCP clients make images and videos for your projects with the AI subscriptions you already pay for. No API keys. No second bill. Uses your own ChatGPT, Google AI Pro, Higgsfield and Grok accounts.

Intro video • Features • Quick start • Connect accounts • Installation • Tools • CLI commands • Troubleshooting

Watch the 72 second intro

https://github.com/user-attachments/assets/c96b3ad1-4795-4c54-91e7-95d3cd4627d3

A short narrated tour, 1080p: the problem it solves, the server making a real image from a prompt, and how to install it. The video itself was made with our own intro video tool and its synthetic NexaForge narrator. Download the MP4 (9.4 MB).

Related MCP server: img-vid-mcp

Features

• Images and videos from your own plans: GPT Image (ChatGPT via Codex), Nano Banana and Veo 3.1 (Google Flow), Gemini, Higgsfield (Seedance, Kling and more), Grok

• No API keys at all: you sign in once, by hand, in a private Chrome window. API-key variables are stripped before any tool runs

• Understands your project: reads what it does, its colours, fonts, 3D or animation libraries and existing images, then matches the style (luxury, premium, playful, corporate, modest)

• Premium by default: writes a real creative brief for every asset, with named looks like neon-glass (the style of the graphics on this page)

• Built-in media studio (new in 0.2): background removal, video editing for Reels and Shorts, and every social size from one image, all running locally

• Saves straight into your project: full quality, no Download folder, no Save dialog, never outside the project

• Shrinks for the web without visible loss: WebP and H.264, checked with an SSIM quality score

• Credit-safe: shows the cost first, hard max_credits cap, one job at a time

• Add any site with a small JSON file, no code

What's new in 1.x

Made the way the server now recommends: the background came from design_brief and image_generate (Codex on a ChatGPT plan), every word and logo was set on top in HTML instead of being drawn by the model, then composition_check and media_optimize checked and shrank it. None of the graphics on this page show another company's logo.

The media studio (since 0.2)

Try:

Remove the background from public/products/lamp.jpg.
Turn this 16:9 clip into a 9:16 Reel with our logo top-right and these subtitles.
Make every social size from public/og.png.

Quick start

git clone https://github.com/farhan6667/no-api-media-mcp
cd no-api-media-mcp
npm ci
node dist/src/index.js setup

npm ci installs the exact dependency versions in package-lock.json and builds the server.

setup checks your machine, opens a Chrome window where you sign in to ChatGPT and Google, and prints the line to add to your AI client. Then just ask your agent for what you need.

No prompt writing needed

You don't write image prompts. When your AI client connects, this server hands it a full workflow, and the agent follows it on its own:

1. Reads your project: what it does (from the README or site), colours, fonts, 3D or animation libraries, existing images

2. Writes a real creative brief: what the project is, where the image goes, who sees it, what it must achieve, and a premium quality bar. This is the part that makes results look high-end instead of generic

3. Generates with the right tool: GPT Image (ChatGPT) for designs with text, Google Flow for photo and 3D scenes

4. Checks every result against a design checklist (spelling, brand fit, no third-party logos, crop) and retries with a corrected brief

5. Shrinks and places the winner in your project, with real alt text

Try one of these

Copy any of these into Claude Code, Cursor or Codex, inside your project:

Look at this project and tell me where images or video would make the site better, then make them and put them in.

Make a premium 16:9 hero image for the landing page that matches our brand.

Create a 1200x630 social preview image for this repo, with the project name spelled exactly.

Make a premium infographic that explains what this project does, for the top of the README.

Make an 8 second calm background video for the hero, quote the credits first.

That's the whole interface. Say "premium", "luxury", "minimal" or "playful" if you want a specific level, otherwise it aims for premium.

Made with this MCP

Both of these came out of the prompts above, through Google Flow on an AI Pro plan, 0 credits, no API key:

At a glance

It designs for who's actually looking

Tell design_brief who sees the asset (context.audience, plain English: "CISOs", "developers", "the open source community", "security analysts", "the general public") and it recognises a few real audiences and changes the art direction for them: tone, a default style tier and a palette when you haven't set your own, plus one honest critique question asked from that reader's seat.

Don't invent an audience it doesn't recognise, say so plainly and it's simply left unset.

It sets the text and logo itself, then delivers the final file

Generated art is only half a post. Register the brands you publish under once:

brand_profiles(action="set", id="sfa", name="SFA", publisher="person", logo_dark="/path/sfa-logo.webp")

Then compose sets the headline and that brand's real logo over the art. It measures the picture to find the calmest corner, keeps the text out of the busy part, picks the text colour and a soft scrim for at least 4.5:1 contrast, and never adds a glow. deliver turns the result into the file a platform wants:

compose(input_path="art.png", headline="Shadow AI", brand_id="sfa", platform="linkedin-feed")
deliver(input_path="art-composed-linkedin-feed.png", platform="linkedin-feed")

That gives an exact 1080x1350 JPEG, as small as it can be without a visible change, metadata stripped, under LinkedIn's limit, and checked for an empty band.

It measures its own work, and remembers your corrections

design_audit doesn't only take the model's word for it: given the file, it measures it, and an empty band at an edge or a washed-out thumbnail caps the score so a weak result can't ship. composition_check does the same on demand and gives a crop that keeps your target shape.

When you say a result was wrong (the crop, the colours, the wrong logo), the assistant records it with design_feedback, and every later brief applies it. A correction can be tied to one brand, so "use my personal mark" never leaks onto company posts.

It handles several sessions at once

Each Claude (or Cursor, or Codex) session runs its own copy of this server, but there's only one signed-in browser profile. By default a second session simply opens its own tab in the same browser, so both jobs run side by side.

This works through a loopback-only Chrome devtools port that's open while the browser is open. On your own computer that's fine. On a shared machine, turn it off:

node dist/src/index.js config set share_browser false

With sharing off, the second session waits in line instead and says so in its progress messages: its place in the queue, how long it has waited, and an estimate once there's enough history (taken from how long recent jobs on your machine really took). It carries on by itself when the browser frees up.

The brief also asks before it guesses now: if it doesn't know where an image will be used or what post it goes with, design_brief returns needs_from_user with those questions, so your assistant asks you once instead of producing the wrong size or the wrong message.

It knows 3D and WebGL sites need different assets

Two more asset types for design_brief: texture (a seamless, tileable surface for a background or a Three.js material) and environment-map (a seamless 2:1 equirectangular panorama for reflections and lighting in a Three.js or WebGL scene). Each gets its own critique checklist instead of the usual photo-realism one, since neither is meant to be looked at directly.

loop_check(input_path=".ai-media/flow/hero-loop.mp4")

Compares a looping video's first and last frame the same way media_optimize checks quality loss, and tells you if the loop will actually look seamless before it ships.

palette_extract(input_path="public/images/hero.jpg", count=6)

Pulls the real dominant colours out of a generated (or any) image and returns a ready :root CSS block, so the site's design system can match what actually got made instead of a guess.

The server also recognises the kind of project it's in on its own now: a GitHub repo, a single social post, a 3D/WebGL site, or an animated landing page, and reaches for the matching asset types without being told each time.

It remembers what actually shipped

design_audit can carry the style (tier and look) a draft used. Once enough drafts of one asset type have shipped under the same tier, design_brief defaults to it on its own next time, instead of always starting from premium. Ask for a specific tier or look and that always wins; this only fills in the gap when you didn't ask.

It compares drafts side by side

contact_sheet(input_paths=["dir1/a.png", "dir1/b.png", "dir1/c.png"], output_path=".ai-media/compare/hero.jpg")

Lays up to 12 images on one grid instead of one at a time. No text is baked into the image; the result lists which grid position holds which file.

It knows what actually keeps platforms from cutting your reach

Meta has said plainly that a plain "AI info" label does not by itself reduce a post's reach. What does get penalised is content that reads as a real photo of an event, person or statement that never happened. design_brief's critique list now asks that question directly for every realistic asset type, and social_sizes covers Facebook, YouTube and X/Twitter cover sizes alongside LinkedIn, Instagram and Pinterest.

It keeps a face or a look consistent

Pass reference_images to image_generate (Codex provider only) and the generated picture is built from your own photos. Use it for a person's portrait, a product, or a style you have already approved. The description says what to keep (for example "the same face, glasses and beard") and what to change (clothes, pose, setting, light).

image_generate(provider="codex", reference_images=["refs/me.png", "approved/portrait-v1.png"], prompt="Keep my face exactly. New navy blazer, glass wall at blue hour, calm expression...")

Up to four images, and each one has to be a real image file inside the project. Only use a photo of someone who has agreed to it, which usually means your own. For a fixed expression or a repeated fix, name the exact fault in the prompt ("brow crease, flat mouth") and pass the previous result as a reference. It is the one place a client can ask for a specific person, so the server never fetches a reference from the web.

It falls back when a provider breaks

Browser-driven providers depend on the site's own pages, and a redesign can break a selector overnight. Pass fallback_providers to image_generate and it tries the next one automatically instead of failing the whole run:

image_generate(provider="flow", fallback_providers=["codex", "gemini"], ...)

The result always lists every attempt and why each one failed, so nothing is silently swallowed. A second local journal (separate from the design one) tracks these outcomes, and when a provider fails the same way three times in a row, the next error says so plainly instead of staying a mystery.

It checks for a new version

Once a day, the server checks this project's GitHub releases. If a newer one is out, it tells you at startup and in status/setup: what changed, and the exact command to update. It never patches itself while running, and it never installs anything unless you turn that on:

The check is a plain, unauthenticated read of this repo's own GitHub releases, with a size cap and a short timeout, and a failed check never breaks a real tool call. Only a plain version number (like 1.4.1) is ever handed to npm.

The package is not published on npm yet, so the setup command and the client config always point at your own copy by its full path, never at npx no-api-media-mcp.

It learns from its own mistakes

Every time the model scores an image with design_audit (and tells it the asset type), the server writes the scores to a small local journal in your no-api-media folder. It is never uploaded and never lives inside a project. The next design_brief for that kind of asset reads it back: any criterion that has fallen short at least twice (for example "not template" or "brand presence") is turned into advice and put into every prompt up front. A mistake that happened twice stops happening a third time.

On top of that, design_brief carries a short list of built-in lessons from real use, such as these:

• Never ask the model to draw a real logo or to spell words. Leave a calm area, then place the real logo file and the text afterwards.

• If the image will be cropped much wider than 16:9, keep every important element in the middle band. Set context.targetAspect (for example 3.2:1 for a profile banner) and the brief adds this guidance for you.

• A single big centred object reads as stock. Ask for an abstract backdrop with open space and let the real brand mark be the focal point.

• A black-background logo dropped on busy art washes out. Put it on a dark glass plate or a feathered mask.

The full list, with the reasons and the provider notes, is in docs/lessons.md.

It keeps your project tidy

Generated drafts land in <project>/.ai-media. Left alone, that folder fills up with rejected attempts. The server now tracks what happened to each draft:

Pass the draft's file to design_audit and media_optimize does the rest on its own. The final image goes to your project's own images folder, so nothing important stays in .ai-media.

Cleanup runs once per session, and you can run it yourself with media_cleanup (a dry run by default, so it first shows what would go and how much space it frees). It is strict on purpose: it only looks inside .ai-media, never touches your final files, skips symbolic links, and ignores anything that is not a media file. If .ai-media itself is a link or junction (a cloned repo could ship one), it refuses to read, write or delete anything there. Turn the automatic part off with node dist/src/index.js config set auto_cleanup false or NOAPI_AUTO_CLEANUP=0.

Connect your accounts

You sign in yourself, once. The server never asks for or sees a password.

Close the whole setup window when you're done. The server can't use the private profile while it is open.

Installation

Requirements: Node 22+, Google Chrome, and ffmpeg for shrinking and editing. Optional: rembg for background removal (uv tool install "rembg[cpu,cli]" or pip install "rembg[cpu,cli]"; the model downloads once on first use).

# all projects
claude mcp add --scope user no-api-media -- node /full/path/to/no-api-media-mcp/dist/src/index.js

# this project only
claude mcp add no-api-media -- node /full/path/to/no-api-media-mcp/dist/src/index.js

Add to your MCP config:

{
  "mcpServers": {
    "no-api-media": {
      "command": "node",
      "args": ["/full/path/to/no-api-media-mcp/dist/src/index.js"]
    }
  }
}

Config file locations:

Restart the app after editing.

.vscode/mcp.json:

{
  "servers": {
    "no-api-media": { "command": "node", "args": ["/full/path/to/no-api-media-mcp/dist/src/index.js"] }
  }
}

codex mcp add no-api-media -- node /full/path/to/no-api-media-mcp/dist/src/index.js

Tools

Where files go

• No output_path: <project>/.ai-media/<provider>/<date>-<prompt>.<ext>, ignored by git, with a manifest.jsonl of every prompt

• With output_path (e.g. public/images/hero.png): must be inside the project and a media file, checked before anything is generated

CLI commands

Without a command, dist/src/index.js runs as the MCP server. That's what your AI client starts.

Troubleshooting

Configuration

All optional.

Add another site without code

Put a file in ~/.no-api-media/providers/<id>.json. Start from examples/providers/template.json. Media is only downloaded from the hosts you list in mediaHosts, and specs are never read from a project folder.

Safety and terms

• Your password never touches this code and the server can't write outside your project. It listens on no network port of its own; the one port is Chrome's loopback-only devtools port while share_browser is on (the default), which you can turn off. The full threat model is in SECURITY.md.

• Web providers drive the consumer sites the way you would by hand. OpenAI's, Google's and xAI's terms restrict automated use of those sites, so using them is your call and your account's risk. The server works at human pace, stops at captchas and sign-in pages, and doesn't hide that it is automation. Official CLIs (Codex, Higgsfield) are the safer routes.

Behaviour notice: metadata is stripped from optimized files

Since 0.3.0, media_optimize removes embedded metadata from the files it writes: EXIF, XMP, text chunks and C2PA content credentials (the signed manifest that ChatGPT, Codex, Google Flow, Gemini and Veo embed, and that platforms such as LinkedIn turn into a "Content credentials" badge). This is what most web image optimizers and CDNs do, and in fact 0.2 already dropped it as a silent side effect of re-encoding. 0.3 makes it explicit: every result lists what was removed, and you can turn it off.

• Your originals are never touched. Generated files in .ai-media/ keep their manifests; only the optimized copy is stripped.

• Turn it off in any of three ways: keep_metadata: true on a call, the environment variable NO_API_MEDIA_KEEP_METADATA=1, or node dist/src/index.js config set strip_ai_metadata false.

• What it does not do. This removes embedded metadata only. It does not remove invisible watermarks such as Google SynthID, which live in the pixels, so platforms and detectors may still identify the AI origin. It is not a way to pass AI images off as real photographs. You are responsible for the rules of the platforms you post to and for any disclosure requirements that apply to you.

• A signed C2PA manifest cannot survive re-encoding anyway (it is bound to the original bytes), so with keep_metadata: true the result is honest about what the encoder lost rather than copying an invalid manifest.

Frequently asked questions

How do I generate images and videos in Claude Code without an API key?

Install this MCP server, then sign in once to ChatGPT, Google (Flow and Gemini), Higgsfield or Grok in the browser window it opens. After that you can ask Claude Code, Cursor or Codex for an image or a video and it uses the plan you already pay for.

Does it see my password?

No. You type your password yourself in a normal browser window. The server never asks for it, types it or stores it. With share_browser off it drives the browser over a pipe only; with it on (the default) Chrome also opens a loopback-only devtools port so several sessions can share the window, which any program on your computer could use while it's open. Turn it off on a shared machine.

Can it bill my API account by accident?

It is built to prevent that. API key variables are removed from every child process, and Codex only runs when its login status shows a ChatGPT login.

Which clients and providers does it work with?

Any MCP client, including Claude Code, Cursor, Codex and Windsurf. Providers are ChatGPT through Codex, Google Flow (Veo and Nano Banana), Gemini, Higgsfield and Grok, plus any other site you describe in a JSON file.

Which assistant model should I use?

The images themselves come from your provider's image model (GPT Image through Codex or ChatGPT, Nano Banana through Flow or Gemini), so picking a different assistant model doesn't change the picture directly. It does change the result: your assistant writes the brief, looks at every draft and scores it honestly, and sets the final text and logo. Use a current top-tier model with good image understanding and tool use; small, fast models tend to skip steps or pass weak drafts. For designs with words, codex renders text best; for photo and 3D scenes, Flow.

What do I need to install to use it?

An MCP client to talk to it, plus whatever sign-in the providers you want need.

• Claude Code: in the terminal, the desktop app, or the Claude Code extension for VS Code. See claude.com/claude-code. Cursor and Windsurf work too.

• Codex as your assistant (optional): the Codex IDE extension for VS Code, Cursor and Windsurf (docs), or the Codex CLI.

• For the codex image provider you need the Codex CLI itself, installed with npm i -g @openai/codex and signed in with codex login using your ChatGPT plan (not an API key). The server calls the CLI, so the editor extension alone isn't enough for this provider.

• For Flow, Gemini and ChatGPT on the web, Google Chrome, and a one-time sign-in in the window accounts_login opens.

Does it leave browser windows open?

No. Each job's tab closes when the job ends, and the window itself closes when your assistant calls browser_done at the end of the work. If that never happens, a safety net closes it after 20 idle minutes (node dist/src/index.js config set browser_idle_minutes 10 changes that). While the work is still going, between rounds of edits, it deliberately stays open so the next step doesn't have to start the browser again.

Does it remove AI watermarks?

No. media_optimize strips embedded metadata (EXIF, XMP and C2PA content credentials) from the files it writes, by default and with a report, and you can switch that off. It never touches invisible watermarks such as SynthID.

Can it do more than generate?

Yes. It removes backgrounds locally, edits video with ffmpeg, exports every social media size and shrinks files for the web. The design_brief and design_audit tools help the model art direct and judge its own results.

Are the providers fine with this?

The browser-driven providers work like you would by hand, and their terms restrict automated use of the consumer sites, so using them is your call. Official command line routes such as Codex and Higgsfield are the safer ones. See Safety and terms.

Does it put text and logos on the image for me?

Yes. compose renders your headline and a registered brand's real logo over the generated art, measuring where the calm area is and checking contrast. The image model itself never draws your logo or your words.

How does it get the right size and format for LinkedIn or X?

deliver takes a platform name (linkedin-feed, x-post, instagram-story, github-social and others) and produces that exact size as an optimized JPEG under the platform's file limit. For a LinkedIn feed post it's 1080x1350; 1200x627 is only for link previews.

I run several Claude sessions. Why did one say the browser is busy?

With share_browser on (the default) sessions work side by side as separate tabs, so you'll mostly see this when sharing is off or the other browser couldn't be joined. The waiting session queues, reports its place and an estimate, and continues on its own.

Can it make assets for a three.js or WebGL site?

Yes: design_brief has an environment-map asset type (a seamless 2:1 panorama for reflections and lighting) and a texture asset type (a seamless tileable surface), on top of the usual hero and background-video. Both get checked for seams instead of photo-realism.

Does Facebook or another platform penalise AI-made images?

Not for being AI-made and disclosed as such; Meta has said that label alone doesn't cut reach. What does get penalised hard is content that reads as a real photo of an event or statement that never happened. design_brief now asks that question for every realistic asset type, and the honest fix is the same as always: make it obviously a graphic or a considered shot, not a faked real moment.

What happens if a provider's page changes and a call breaks?

Pass fallback_providers to image_generate (for example ["codex", "gemini"]) and it tries the next one in order. The result reports every attempt and its failure reason. A local, redacted journal also tracks this over time, so a real pattern (the same provider failing the same way repeatedly) gets called out instead of looking like one bad run.

Can it keep my face the same across pictures?

Yes, with Codex. Pass reference_images with one or more photos of yourself, and say in the prompt what must stay (face, glasses, beard) and what may change. It is meant for your own photos or people who agreed, and results still need a look from you before they go public.

Does it know who the image is for?

If you tell design_brief who sees it (context.audience), it recognises a short list of real audiences (CISOs, developers, the open source community, security analysts, the general public) and adjusts tone, default tier, palette and the critique question accordingly. It won't guess an audience you didn't name.

Will it update itself without telling me?

No. It checks once a day and tells you when a newer version exists, with the exact command to run. For an npm install it only installs the update for you if you explicitly set NOAPI_AUTO_UPDATE=1, and even then only for the next run, never while the server is live. A git checkout is never updated automatically.

Where do the palette and font suggestions in design_brief come from?

From a reviewed UI design skill's colour and typography database, picked by the same topic detection the brief already does. They are only used when you don't supply your own brand colours, and you can always override them.

Does it learn from my past results?

Yes, locally. design_audit writes each score sheet to a small journal on your machine, and design_brief turns the criteria that keep falling short into advice at the top of the next prompt for that asset type. It also carries built-in lessons from real use. Nothing is uploaded.

Will it fill my disk with old drafts?

No. Rejected drafts are removed after 3 days, drafts that already became a final file after 14 days, and untouched ones after 30. It only ever looks inside .ai-media, never touches your final files, and you can switch it off or run it by hand with media_cleanup.

Is it free?

The server is free and open source under Apache-2.0. It uses the AI plans you already have, so it adds no API bill.

Contributing

Fixes for broken selectors and new site specs are the most useful help. See CONTRIBUTING.md.

Contact and support

• Tried it? Tell me how it went, good or bad, with the feedback form. It takes two minutes.

• Bugs, questions, ideas: open a GitHub issue or start a discussion.

• Security problems: use a private security advisory, see SECURITY.md.

• Work with NexaForge (cyber security, vibe coding, web development and IT infrastructure): nexaforge.services@gmail.com or nexaforge.eu.cc/contact.

• If this saved you an API bill, a star on the repo is the easiest way to say thanks.

• Sponsor the work. If your company can help with hosting, tools or funding, the sponsor page explains how. GitHub Sponsors is not available in Pakistan, so it works by email.

License

Apache-2.0. See NOTICE for the attribution and trademark note.

Built by Syed Farhan Ahmed (SFA) at NexaForge Cyber security · Vibe coding · Web development and IT infrastructure

Website · LinkedIn · Portfolio · Email