AdAnvil

An MCP server that lets an AI agent edit vertical 9:16 video ads.

AdAnvil works with Claude Code, Codex, Cursor or any MCP client. It builds ads from a brand's own footage, from a reference ad, or from an idea. The agent judges (which clips, what the text says, whether the result looks right). AdAnvil renders deterministically and checks the result: the same storyboard always gives the same video.

It is not tied to a niche. The recipes and the editing guide speak in general terms (product, benefit, problem, target customer) with examples from several niches: coffee, fashion, fitness, pets, software, home, beauty.

demo

The demo above was made from the CC0 demo pack in examples/demo/ (four Wikimedia Commons clips, a product still and a fictional brand called Ember & Oak). No footage is stored in this repository: the pack is fetched on demand.

What it does

• Reference decoding: shots, pacing, burnt-in text and speech, plus a shot-by-shot breakdown the agent writes and compares to your edit.
• Footage library: indexing, contact sheets and mosaics so the agent can look at every clip before using it, and the speech of every video and audio file (word timestamps, searchable as soon as each file is done) so a talking creator can be cut on the pauses.
• 9 recipes (YAML) for common ad shapes: text hook, hook plus product, before and after, customer diary, comment reply, meme text, voice-over ad, long remake, talking-head UGC.
• 7 caption styles (pill, meme, outline, outline_soft, box, label, comment) with emoji support.
• Removal of burnt-in text from reference footage.
• HDR (HLG and PQ) to SDR tone-mapping.
• Music matching and loudness control.
• Optional open-source voice-over (Chatterbox) with captions synced to the spoken words.
• Talking-head UGC: best take of each line across several recordings, cuts on the pauses without hesitations, the creator's own sound kept (with B-roll laid over the voice), captions built from the speech.
• Visual devices: a before/after reveal (a divider sweeps one picture away to show the other), split screens, picture cards (a photo or a screenshot that comes in over the video), video insets (picture-in-picture), stickers and arrows, a progress bar and punch-in zooms.
• Sound design: a pack of about 18 sound effects (CC0 or synthesised, never stored in the repository), sfx overlays and audio.sfx_auto at the cuts and captions, and cuts that snap to the beat of the music.
• End cards composed from a real product photo.
• Quality control with images: automatic checks, contact sheet of the result, full-resolution frames.
• Detached background jobs for indexing, voice and rendering, so long work survives a dropped connection.

Install

Pick one path. The first is the simplest.

1. Claude Desktop extension (one click)

Download adanvil-<version>.mcpb from the latest release, double-click it, and click Install in Claude Desktop. It asks for a projects folder and, optionally, a Freesound key (stored securely). You still need ffmpeg 6.1 or newer on your computer (brew install ffmpeg, winget install Gyan.FFmpeg, sudo apt install ffmpeg); if it is missing, the extension says so with the command for your system.

2. Docker (no Python, no ffmpeg to install)

git clone https://github.com/maxbuildog/adanvil && cd adanvil
docker build -t adanvil .

The image has ffmpeg 6.1, runs as a non-root user and speaks MCP over stdio. Mount the folder that holds your projects at /projects; the agent must give tool calls paths under /projects (for example /projects/my-ad).

# Claude Code
claude mcp add adanvil -- docker run -i --rm -v "$PWD:/projects" adanvil

On Linux add --user "$(id -u):$(id -g)" so the files written to the folder belong to you.

Codex, in ~/.codex/config.toml:

[mcp_servers.adanvil]
command = "docker"
args = ["run", "-i", "--rm", "-v", "/home/you/ads:/projects", "adanvil"]
tool_timeout_sec = 900
default_tools_approval_mode = "approve"  # needed with approval_policy = "never" (codex exec)

Cursor, in .cursor/mcp.json:

{
  "mcpServers": {
    "adanvil": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "-v", "/home/you/ads:/projects", "adanvil"]
    }
  }
}

The image holds the base package only. Transcription is available with docker build --build-arg EXTRAS=transcribe -t adanvil:transcribe .. The voice, vision and music-generation environments (adanvil voice-setup, vision_setup, music-setup) are not included: they need gigabytes of models, so use the uvx path for them.

3. uvx

Requirements: Python 3.11 or newer, uv (it provides uvx), and ffmpeg 6.1 or newer, built with libzimg, on your PATH (check with ffmpeg -version; if it lives elsewhere, set ADANVIL_FFMPEG and ADANVIL_FFPROBE to the full paths).

# 1. Check the install (the first run downloads the package)
uvx --from git+https://github.com/maxbuildog/adanvil adanvil --version

# 2. Register the server in Claude Code
claude mcp add adanvil -- uvx --from git+https://github.com/maxbuildog/adanvil adanvil-mcp

# 3. Then ask your agent:
#    "Make a 9:16 ad from the footage in ./my-project"

The package is named adanvil. It installs two commands: adanvil (command line) and adanvil-mcp (the MCP server over stdio). uvx adanvil-mcp alone does not work, because uvx looks for a package with that name: always add --from. Once the package is on PyPI you can use --from adanvil instead of the Git URL.

The agent creates the project folder and tells you where to drop your files: footage in footage/, the reference ad in reference/, music in music/, logo and end card in brand/. It learns the method from the editing_guide tool and from AGENTS.md.

Codex, in ~/.codex/config.toml:

[mcp_servers.adanvil]
command = "uvx"
args = ["--from", "git+https://github.com/maxbuildog/adanvil", "adanvil-mcp"]
tool_timeout_sec = 900                   # indexing long footage can take minutes
default_tools_approval_mode = "approve"  # needed with approval_policy = "never" (codex exec)

Without default_tools_approval_mode, a non-interactive codex exec run rejects every AdAnvil tool call with "MCP tool call requires approval".

Cursor, in .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "adanvil": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/maxbuildog/adanvil", "adanvil-mcp"]
    }
  }
}

Any other MCP client (generic stdio entry):

{
  "mcpServers": {
    "adanvil": {
      "type": "stdio",
      "command": "uvx",
      "args": ["--from", "git+https://github.com/maxbuildog/adanvil", "adanvil-mcp"],
      "env": {}
    }
  }
}

Restart the client after editing its config, then ask the agent to call recipe_list. Nine recipes means the server is connected.

Optional: reference transcription needs the transcribe extra (faster-whisper). Use --from "adanvil[transcribe] @ git+https://github.com/maxbuildog/adanvil" in place of the plain --from value above.

Try it with the demo pack

The demo needs a checkout, because the fetch script lives in the repository:

git clone https://github.com/maxbuildog/adanvil
cd adanvil
uv sync
uv run python examples/demo/fetch_demo.py ./demo-project
uv run adanvil index ./demo-project --source footage

Then register the MCP server and ask your agent to make an ad from ./demo-project (give it the absolute path), or follow the command line walkthrough in examples/demo/README.md.

How it works

Everything lives in one project folder: footage/, brand/, music/, reference/, renders/, cache/, plus adanvil.json, brand.yaml and catalog.json. Every tool takes the absolute project_dir; every other path is relative to it and cannot escape it.

1. The agent indexes the footage, looks at the contact sheets, and optionally decodes a reference ad.
2. It picks a recipe and writes a storyboard (JSON): shots, captions, audio.
3. storyboard_validate checks the storyboard against the recipe rules, then render_start renders it in the background.
4. qc_run checks the file and returns control images that the agent looks at before delivering; ad_score scores its creative quality (aim for 80 or more).

Details: docs/ARCHITECTURE.md for the pipeline, docs/EDITING_GUIDE.md for the editing doctrine, and AGENTS.md for the step-by-step guide an agent follows (same content as skill/SKILL.md, the Claude Code skill version).

Choosing a model

The result depends on the model that drives AdAnvil. Reference analysis and shot choice are judgement work: use the strongest model you can (for example Claude Opus) for them. Smaller models execute the steps well, but they tend to copy the surface of a reference (same words, same shots) instead of rebuilding its logic for your brand.

Tools (43)

CLI commands (adanvil <command> --help): mcp, voice-setup, vision-setup, init, cat, onboard, index, decode, recipes, validate, render, voice, qc, preview-caption, end-card, sfx-fetch. Add --json for machine-readable output.

Smart library (optional)

Finding the right shot is the slowest part of an edit. adanvil vision-setup (or the vision_setup tool) installs, once, a separate environment (about 2.5 GB, CPU only, no GPU needed) with a small multilingual image and text model (SigLIP 2, Apache-2.0) and a face detector (YuNet, MIT); the licences are in THIRD_PARTY.md. Nothing else changes without it, and library_index then says vision: unavailable.

• Index with vision. library_index also embeds the frames of each contact sheet and looks for faces (cache/vision/<file key>.npz), and gives every entry auto_tags (a fixed label list such as person, bathroom, hands, smiling, text on screen), kept apart from your own tags.
• Search by sentence. library_search(query="a woman with a sad face touching her jaw"), in any language, ranks shots across the whole library, with the usual filters applied first, and returns the best ones with one numbered mosaic: no sheet to open.
• Product versions from photos. brand.yaml lists product_versions: [{name: current, photos: [...]}, {name: old, photos: [...]}]; indexing guesses which version each shot shows, and library_search(product_version="current") uses the guess unless a human annotation exists (a human annotation always wins). Give photos, not words.
• Face-aware framing. focus: "face" keeps the main face centred at face_y (0.38 of the frame by default) on clips, images, split panels and reveal sides; a reveal also puts both faces at the same place and size by default (align_faces). compare_pair uses the face detector too.

Voice-over (optional)

AdAnvil can add a generated voice-over with Chatterbox, an open-source text-to-speech model, with captions synced to the spoken words.

• Run adanvil voice-setup once. It creates a separate Python environment (in the AdAnvil cache folder) and downloads about 3 GB of models. adanvil voice-setup --engine kokoro adds the fast engine (below) to the same environment.
• Generation runs on the CPU. One worker process per job loads each model once and serves every step (speech, alignment, check). Measured on an idle 8-core server (French, 3 to 9 lines of about 5 seconds): the Chatterbox voice takes about 25 to 35 seconds per line, so 9 lines take about 5.5 minutes (it took 10.5 before V2.6); a busy machine is slower, count up to 5 minutes per sentence.
• Two engines. chatterbox is the default and the recommended voice for ads (the owner listened to both and kept it): it clones the reference voice, so it also fits a brand that has a voice to match. kokoro (Apache-2.0) has fixed voices and no cloning ({"engine": "kokoro", "voice": "ff_siwis"}, French ff_siwis, English af_heart by default, bf_emma and others for British English): a fast option for drafts and previews (check the script, the timing and the captions before paying for the real voice), about 25 seconds for 9 lines plus 50 seconds of alignment and check on the same server. It speaks French and English only.
• The pronunciation check reuses the alignment transcript when every word is confident (check_confidence, default 0.6) and the line reads as written; only doubtful lines are transcribed again by the check_model. Set check_confidence to null (or ADANVIL_VOICE_CHECK_CONFIDENCE=off) to check every line with the check_model.
• The voice speaks for the brand, as a narrator. It must not pose as a fake customer.
• The voice language is the language of the voice spec. When the spec does not set it, AdAnvil uses the language of the project's brand.yaml (fr-FR counts as fr), and en when the brand has none.
• Chatterbox embeds a Perth watermark in its audio. AdAnvil keeps it. Do not remove it.
• Use only a reference voice you have the right to use.

Talking-head UGC (optional)

A creator talking to camera is cut from the speech in the footage. library_index transcribes the clips and audio files that have sound (word timestamps, in the same job, once per file, each file stored as soon as it is done), with faster-whisper when it is installed (pip install 'adanvil[transcribe]') or else the voice environment (adanvil voice-setup); without either, indexing still succeeds and the entries say so. Indexing uses a fast model to make the speech searchable; speech_refine transcribes the clips you keep again with a stronger one, and the brand name, the product name and speech.vocabulary of brand.yaml are given to the transcriber. A voice recorded in a separate audio file is found by name (separate_audio_candidates on a video with silent sound) and laid with audio_from; syncing is up to the agent, on a preview. Then speech_takes picks the best take of each line, speech_view gives the cuts that keep the speech and drop pauses and hesitations, and a storyboard with audio.clips, audio_from (B-roll over the voice) and a speech_captions overlay renders the ad. The ugc_talking recipe puts it together; the guides explain each step.

• ADANVIL_SPEECH_MODEL sets the indexing model (default base: fast, good enough to search). speech_refine uses medium unless told otherwise. The language is the language of brand.yaml, else auto-detected.
• ADANVIL_SPEECH=off stops library_index from transcribing on its own (transcribe="auto"); transcribe=true still does.
• The creator's words are claims: the validator warns banned_word_spoken when a kept word is on the brand's banned list, and cut_mid_word when a cut falls inside a word (not when the voice goes on across the cut). The fillers that are cut (euh, um, uh, er, genre...) can be replaced in brand.yaml under speech, which also takes a vocabulary list of words to spell right. A misheard word can be fixed in the captions with speech_captions.corrections, close spellings only, and every correction is listed for review (speech_corrections).

Visual devices (optional)

A reveal segment (a divider line sweeps a before picture away to show the after picture, from two real pictures or two clips, with labels that follow their side) for a before/after the viewer watches happen, a split segment (side by side or top and bottom, with labels such as FROM THIS and TO THIS) for comparisons, an image overlay with animation, frame and rotate (a photo or a review screenshot that comes in over the video as a card), a video overlay (picture-in-picture, with a rect, rounded or circle shape, border and shadow), sticker overlays (arrow, hand-drawn circle and underline, badge, emoji, image, with a pop or draw animation), a progress bar and a punch (instant zoom) on a clip. Use one or two per ad. Fields and rules: AGENTS.md and the editing guide.

Sound design (optional)

Sounds go over the audio bed of an ad: a whoosh on a cut, a pop on a caption, a ding on the price, a riser before the reveal.

• adanvil sfx-fetch <project_dir> (or sfx_list with fetch=true) builds a pack of about 18 short sounds in the project's sfx/ folder. Each one is downloaded from a pinned CC0 source (checked by SHA-1) or synthesised in code, then trimmed and set to one level. Nothing is stored in this repository; sources and licences are listed in examples/sfx/README.md and written to sfx/LICENCES.txt. Your own files in sfx/ work too.
• {"type": "sfx", "source": "sfx/pop.wav", "at": 3.2, "volume": 0.8} places a sound; audio.sfx_auto places them at the cuts and at the captions. Sounds never clip and never run past the end.
• music_beats gives the beats of a track as it will play, and storyboard_snap moves the cuts onto them (the validator reports how many cuts are already on a beat).

Performance settings

• ADANVIL_WORKERS: how many files, thumbnails or segments are processed at once. Default max(1, min(4, cpus // 2)), and at most 2 when the machine has under 12 GB of RAM.
• ADANVIL_FFMPEG_THREADS: threads of each ffmpeg decoder or encoder. Default max(1, cpus // workers).
• Voice: ADANVIL_VOICE_CHECK_MODEL sets the Whisper model of the pronunciation check (unless the voice spec sets check_model). ADANVIL_VOICE_THREADS sets the torch and Whisper threads (default: usable cores). ADANVIL_VOICE_CHECK_CONFIDENCE is the word confidence from 0 to 1 above which the alignment transcript replaces the check model (off to always run the check model; the spec's check_confidence wins). ADANVIL_VOICE_PERSISTENT=0 goes back to one process per step instead of one worker per job.
• Speech: ADANVIL_SPEECH_MODEL (default base), ADANVIL_SPEECH=off (no automatic transcription), ADANVIL_SPEECH_PROMPT (an optional text that biases the transcriber, for example towards keeping hesitations).
• Other: ADANVIL_CACHE moves the cache folder; ADANVIL_FFMPEG and ADANVIL_FFPROBE point to the binaries.
• Benchmark: uv run python scripts/bench.py --files 12 --seconds 4.

Responsible use

AdAnvil edits ads; you answer for what they say and show.

• Use real comments and real customer words only. The tool never invents an offer, a price, a guarantee or a claim: the human confirms them.
• Get the consent of the people who appear in your footage.
• Check the licences of the music and the footage you use (the demo pack is CC0, your own files may not be).
• Make sure every claim in the ad is true and allowed in your market.
• Do not clone a voice without the right to do so, and do not present a generated voice as a real customer.

Contributing

See CONTRIBUTING.md.

Licence and status

MIT, see LICENSE. Fonts, emoji and other third-party material are listed in THIRD_PARTY.md.

Status: alpha, version 0.1.0. Interfaces can still change. See CHANGELOG.md.