Skip to main content
Glama

resolve-mcp-kit

Two tools in one folder.

  1. A Resolve MCP server. It lets an AI agent (Claude Code, Claude Desktop, opencode, any MCP client) see and drive DaVinci Resolve on your computer: read projects, import clips, build timelines, add titles, render.

  2. site-video. You give it a website link. It explores the site, has a language model write a script, records the real site in a browser, adds captions and an optional voiceover, and edits a promo video. It works with or without Resolve. Without Resolve, ffmpeg joins the clips.

You can use either one alone. Follow the sections in order for a first video. Skip to With Resolve for the MCP server.

What you need

Part

Windows

macOS

Linux

site-video, ffmpeg editor (no Resolve)

works

should work, not run by the author

works (run on Ubuntu 24.04, Python 3.12)

site-video, Resolve editor

built and used here (Resolve Studio 21.1)

untested

untested

Resolve MCP server

built and used here

untested

untested

Sign in as me (a window on the queue's computer)

Chrome or Edge

Chrome or Edge, untested

Chrome, Chromium or Edge, untested

Browser extension sign-in

Chrome or Edge

Chrome or Edge

Chrome or Edge

  • Python 3.10 or newer. 3.12 is a safe choice. The Resolve scripting module also has to match the Python that Resolve supports.

  • Disk space. About 1 GB for the install (the packages, plus Chromium and its helpers, which Playwright downloads once). Each video job keeps its recorded clips, voice files and the final MP4. A 30 second video took 18 MB on the ffmpeg path. Long videos, the clips deliverable and Resolve renders take much more. Delete jobs you no longer need.

  • DaVinci Resolve Studio (paid) for anything that touches Resolve. Blackmagic's free edition does not allow external scripting in most versions, so expect "Could not attach to Resolve" there. Resolve must run on the same computer as the server, with a project open.

  • A language model that writes the script. This is the one thing site-video cannot do without. A free local one (Ollama) works, no API key needed. See Model setup.

  • Optional: a voice (text to speech) for narration, a vision model, speech to text, Chrome or Edge for sign-in. Without a voice the video has captions and no narration.

Related MCP server: io.github.Airta-Admin/davinc-resolve-local-bridge-mcp

Install

Pick your system. Run the commands in the folder you cloned.

Windows (cmd.exe or PowerShell, Python from python.org with "Add to PATH" ticked):

git clone https://github.com/Jaron-Wilson/resolve-mcp-kit.git
cd resolve-mcp-kit
py -3 -m venv .venv
.venv\Scripts\python.exe -m pip install -r requirements.txt -r requirements-video.txt
.venv\Scripts\python.exe -m playwright install chromium

In the commands below, python means the venv's Python. Either activate it once per window (.venv\Scripts\activate in cmd.exe, .venv\Scripts\Activate.ps1 in PowerShell) or type .venv\Scripts\python.exe instead of python.

macOS or Linux:

git clone https://github.com/Jaron-Wilson/resolve-mcp-kit.git
cd resolve-mcp-kit
python3 -m venv .venv
. .venv/bin/activate
python -m pip install -r requirements.txt -r requirements-video.txt
python -m playwright install chromium

On Debian or Ubuntu, python3 -m venv may need sudo apt install python3-venv first. On a bare Linux server, Chromium may need system libraries: sudo .venv/bin/python -m playwright install-deps chromium.

The start scripts do the same install for you. start-resolve-mcp.cmd (Windows) and ./start-resolve-mcp.sh (macOS and Linux) create .venv, install both requirements files, download Chromium, then start the MCP server. They do this once. Set RESOLVE_MCP_SKIP_VIDEO=1 first to skip the video packages and the Chromium download if you only want the Resolve MCP server. Set RESOLVE_MCP_REINSTALL=1 once after you update this folder.

Check the install (optional, about 15 seconds). Run the tests. They need no model, no Resolve and no network:

python -m pip install pytest
python -m pytest tests

Every test should pass (the last line reads like 81 passed).

ffmpeg needs no install: the imageio-ffmpeg package ships a copy. Set FFMPEG to the path of your own ffmpeg to use that instead.

Model setup

site-video needs one chat model to write the script. It is any server with an OpenAI-style /chat/completions API. Everything is set with environment variables, in the same window where you start site-video.

Variable

Default

What it is

SITE_VIDEO_LLM_URL

http://127.0.0.1:11434/v1

Base URL of the API, ending in /v1. The default is a local Ollama.

SITE_VIDEO_LLM_MODEL

qwen3:8b

The model id the server lists.

SITE_VIDEO_LLM_KEY

empty

API key, sent as a bearer token. Leave it empty for local servers.

Set a variable like this:

bash, zsh:    export SITE_VIDEO_LLM_MODEL=qwen3:8b
cmd.exe:      set SITE_VIDEO_LLM_MODEL=qwen3:8b
PowerShell:   $env:SITE_VIDEO_LLM_MODEL = "qwen3:8b"

Free and local: Ollama (no API key)

  1. Install Ollama from https://ollama.com and start it. It listens on 127.0.0.1:11434.

  2. Pull the default model (a download of about 5 GB): ollama pull qwen3:8b

  3. Check it: curl http://127.0.0.1:11434/v1/models should list qwen3:8b.

With Ollama running you need no settings at all. A bigger model writes better scripts. Pick one with SITE_VIDEO_LLM_MODEL, or per video with --model. Thinking models work: the reasoning is cut out of the answer.

Any other OpenAI-compatible service

SITE_VIDEO_LLM_URL=https://api.openai.com/v1      SITE_VIDEO_LLM_KEY=sk-...   SITE_VIDEO_LLM_MODEL=<a chat model id from your account>
SITE_VIDEO_LLM_URL=http://127.0.0.1:1234/v1       (LM Studio; set the model to the one you loaded)
SITE_VIDEO_LLM_URL=https://openrouter.ai/api/v1   SITE_VIDEO_LLM_KEY=sk-or-...  SITE_VIDEO_LLM_MODEL=<an id from their list>

vLLM, llama.cpp's server and others work the same way: give their /v1 URL.

The queue page lists the models the endpoint reports (GET /models). If your server does not list them, type the id.

Voice (optional)

Choose the voice per video with --voice (or the page's Voice menu). SITE_VIDEO_DEFAULT_VOICE sets the default.

Voice

Needs

Notes

none

nothing

No narration. This is the default and needs no key.

openai[:voice]

an API key in SITE_VIDEO_TTS_KEY (or OPENAI_API_KEY)

Any /audio/speech API. SITE_VIDEO_TTS_URL (default https://api.openai.com/v1), SITE_VIDEO_TTS_MODEL (default tts-1), voice default alloy.

openai[:voice] with a local server

the server's URL

For example a Kokoro-FastAPI server: SITE_VIDEO_TTS_URL=http://127.0.0.1:8880/v1, SITE_VIDEO_TTS_MODEL=kokoro, --voice openai:af_heart. No key needed. Not run by the author.

elevenlabs:<voice id>

ELEVENLABS_API_KEY

SITE_VIDEO_ELEVENLABS_MODEL (default eleven_multilingual_v2).

odysseus

SITE_VIDEO_ODYSSEUS_URL, SITE_VIDEO_ODYSSEUS_TOKEN

Only if you run an Odysseus server. Skip it otherwise.

The queue page offers only the voices that are configured. SITE_VIDEO_OPENAI_VOICE and SITE_VIDEO_ELEVENLABS_VOICE set the voice name or id the page fills in.

Vision and speech to text (optional)

These matter only for your own videos and for the automatic check of the finished scenes.

Variable

Default

What it does

SITE_VIDEO_VISION_MODEL

empty

A vision model on the same SITE_VIDEO_LLM_URL endpoint, for example qwen2.5vl:7b. Describes frames of your videos, and turns on the check that each scene shows what its caption says.

SITE_VIDEO_REVIEW

1

Set to 0 to turn that check off. It only runs when a vision model is set.

SITE_VIDEO_STT

empty

openai for speech to text from your videos. Then SITE_VIDEO_STT_URL (default https://api.openai.com/v1), SITE_VIDEO_STT_KEY (or OPENAI_API_KEY), SITE_VIDEO_STT_MODEL (default whisper-1). A local faster-whisper server works. odysseus is the other option.

Quick start: your first video, no Resolve

You need the install above and a running model. Everything below works on Windows, macOS and Linux.

  1. Pick a site you are allowed to record. Your own site is the best choice. Do not record sites you have no right to show.

  2. Run one command:

    python -m site_video make https://example.com --length 30s --editor ffmpeg --out my-first-video

    It explores the site, writes the script, records, and edits. It prints progress lines such as [ 10%] writing the script. A 30 second video takes a few minutes. The model is the slow part.

  3. When it ends it prints a JSON summary. The video line is the full path: it is in my-first-video/deliverables/ and is named after the video's title. Open it with any player. The folder also has plan.json (the script), log.txt and crawl.json.

Without --out, the job goes to ~/site-videos/jobs/<time>-cli/ (SITE_VIDEO_HOME moves it).

Want narration? Add --voice openai:alloy after setting up a voice. Want a longer video, or a say in what it shows? Use --length 2min --notes "show pricing, end on sign-up".

The same thing from a web page is in the next section. Your own videos, changing a finished video and the other options are in More ways to use site-video.

The queue and its web page

The queue runs several videos at once and gives you a page to use it.

python -m site_video serve

It prints site-video on http://127.0.0.1:8940. Open http://127.0.0.1:8940/ in a browser. Paste a link, pick a length, press Make the video. You can watch a live picture of what the browser is doing, then play or download the result.

From a terminal, in a second window:

python -m site_video add https://example.com --length 30s --editor ffmpeg --wait
python -m site_video list
python -m site_video get <id> -o video.mp4

Options of serve: --host (default 127.0.0.1), --port (default 8940), --parallel (videos prepared at once, default 2), --home (where jobs and uploads live, default ~/site-videos).

Binding safely

  • The default is 127.0.0.1: only this computer can reach it. Keep it that way if you can.

  • The queue can make your computer open any web address you give it, upload files to it, and (if you use sign-in) hold logins. Treat access to it like access to your account.

  • Before you bind to another address (--host 0.0.0.0 or a LAN or VPN address), set a token before you start it:

    export SITE_VIDEO_TOKEN=a-long-random-string        # cmd.exe: set SITE_VIDEO_TOKEN=...
    python -m site_video serve --host 100.x.y.z

    Every request then needs Authorization: Bearer <token>. In a browser, open http://100.x.y.z:8940/?token=<token> once. It sets a cookie. The command line tools read SITE_VIDEO_TOKEN from your environment.

  • The token travels in clear text over plain HTTP. Use a VPN you trust, or put an HTTPS proxy in front. Never forward the port on your router.

  • Without a token, anyone who can reach the port can use the queue.

With Resolve: the MCP server

This part needs DaVinci Resolve Studio on the same computer. The author built and used it on Windows only.

1. Allow scripting in Resolve. Open Resolve and a project. Go to Preferences > System > General and set External scripting using to Local. Restart Resolve if it asks. Keep the project open: the API attaches to the open Resolve.

2. Start the server.

Windows: double-click start-resolve-mcp.cmd. macOS or Linux: chmod +x start-resolve-mcp.sh once, then ./start-resolve-mcp.sh.

It prints resolve-mcp on http://127.0.0.1:8930/sse (sse, auth=none, write=on). Leave the window open. Resolve's API only works in your logged-in desktop session, so do not run it as a Windows service.

3. Connect a client. Two styles. The network style connects by URL to the running server. The stdio style makes the client start the server itself.

Claude Code (network):

claude mcp add resolve --transport sse http://127.0.0.1:8930/sse

With a token (see below) add --header "Authorization: Bearer YOUR_TOKEN". Or use stdio, with your real path:

claude mcp add resolve -e RESOLVE_MCP_TRANSPORT=stdio -e RESOLVE_MCP_ALLOW_WRITE=1 -- /path/to/resolve-mcp-kit/.venv/bin/python /path/to/resolve-mcp-kit/resolve_mcp_server.py

On Windows use C:\path\to\resolve-mcp-kit\.venv\Scripts\python.exe and C:\path\to\resolve-mcp-kit\resolve_mcp_server.py.

Claude Desktop (stdio): open Settings > Developer > Edit Config. That opens claude_desktop_config.json. Merge this in, fix the two paths, save, and restart Claude Desktop. This is examples/claude_desktop_config.json.

{
  "mcpServers": {
    "resolve": {
      "command": "C:\\path\\to\\resolve-mcp-kit\\.venv\\Scripts\\python.exe",
      "args": ["C:\\path\\to\\resolve-mcp-kit\\resolve_mcp_server.py"],
      "env": { "RESOLVE_MCP_TRANSPORT": "stdio", "RESOLVE_MCP_ALLOW_WRITE": "1" }
    }
  }
}

On macOS use "/path/to/resolve-mcp-kit/.venv/bin/python" and "/path/to/resolve-mcp-kit/resolve_mcp_server.py". With stdio you do not need the start script window, but the venv must already exist (run the install above, or run the start script once).

opencode: merge the mcp block from examples/opencode.local.json (stdio, edit the two paths) or examples/opencode.remote.json (URL; it reads the token from RESOLVE_MCP_TOKEN in the environment, delete its headers block if you set none). Then run opencode mcp list.

Other clients: point them at http://127.0.0.1:8930/sse, or start with RESOLVE_MCP_TRANSPORT=streamable-http and use http://127.0.0.1:8930/mcp.

4. First command. In your client, ask: "Call resolve_status." You should see connected: true, the Resolve version and your project. If you see connected: false, the reason says what to fix. See Troubleshooting. Then try "List my projects" and "What is on my timeline?".

Skills. skills/davinci-resolve-editing and skills/website-promo-video teach an agent the editing and video workflows. Copy a folder to ~/.claude/skills/ (Claude Code) or ~/.config/opencode/skills/ (opencode; on Windows %USERPROFILE%\.config\opencode\skills\). A copy is also in .claude/skills/, so opening Claude Code in this folder picks them up.

What the MCP server can do

Always on (read-only, plus changing the visible page):

Tool

Purpose

resolve_status

Is Resolve reachable; product, version, page, project, timeline

list_projects

Projects in the current database folder

list_timelines, timeline_info

Timelines and their frame range, tracks, fps

list_media_pool

Clips in a bin: name, duration, fps, resolution

set_page

Switch to media, cut, edit, fusion, color, fairlight, deliver

project_settings

Read project settings such as frame rate and resolution

Editing tools, only when RESOLVE_MCP_ALLOW_WRITE=1 (the start scripts set it; nothing here deletes anything):

Tool

Purpose

import_media

Import files into the media pool, optionally into a bin

create_timeline

New empty timeline, made current

append_clips

Append clips in order, optionally trimmed (cuts)

add_title

Insert a Text+ title and set its text

add_marker

Timeline marker

save_project

Save

render_formats_and_codecs, start_render, render_status

Queue, start and watch a render

make_site_video, edit_site_video, site_video_status, list_site_videos

Use a site-video queue (SITE_VIDEO_SERVER, default http://127.0.0.1:8940)

MCP server settings

Variable

Default

Meaning

RESOLVE_MCP_TRANSPORT

sse

stdio, sse, or streamable-http

RESOLVE_MCP_HOST

127.0.0.1

Bind address for sse and http

RESOLVE_MCP_PORT

8930

Port for sse and http

RESOLVE_MCP_TOKEN

unset

If set, requests need Authorization: Bearer <token>

RESOLVE_MCP_ALLOW_WRITE

0 (the start scripts set 1)

Register the editing and render tools

RESOLVE_SCRIPT_API, RESOLVE_SCRIPT_LIB

per system, below

Where Resolve's scripting module lives

sse serves /sse. streamable-http serves /mcp.

The server finds Resolve's scripting module here. Set the two variables if your install is elsewhere.

System

RESOLVE_SCRIPT_API

RESOLVE_SCRIPT_LIB

Windows

%PROGRAMDATA%\Blackmagic Design\DaVinci Resolve\Support\Developer\Scripting

C:\Program Files\Blackmagic Design\DaVinci Resolve\fusionscript.dll

macOS

/Library/Application Support/Blackmagic Design/DaVinci Resolve/Developer/Scripting

/Applications/DaVinci Resolve/DaVinci Resolve.app/Contents/Libraries/Fusion/fusionscript.so

Linux

/opt/resolve/Developer/Scripting

/opt/resolve/libs/Fusion/fusionscript.so

The Windows paths were used by the author. The macOS and Linux paths come from Blackmagic's documentation and are untried.

Security. This server can open, change and render your projects and write files wherever Resolve can. Keep it on 127.0.0.1. If the client is on another machine, bind to a LAN or VPN address only, set RESOLVE_MCP_TOKEN to a long random string, and never expose the port to the internet. The token is sent in clear text over plain HTTP, so it stops casual access, not someone watching your network.

Making the video in Resolve

With Resolve open, run the queue on the same computer and use --editor resolve (or leave it on auto, which uses Resolve when it answers and ffmpeg when not). Each video becomes its own project in a "Site videos" project folder: track V1 "Picture" with a dissolve on every cut, V2 "Captions", A1 the soundtrack, then a render. You can polish it by hand.

  • SITE_VIDEO_START_RESOLVE = Resolve's program file: if set, the queue opens Resolve when it is closed and closes it again afterward (SITE_VIDEO_START_RESOLVE_ARGS, default -nogui). Run the queue in your desktop session.

  • SITE_VIDEO_RESOLVE_FOLDER = the project folder's name (default Site videos).

  • Resolve 21 opens on its Project Manager screen. While it is up, scripting cannot open or create projects. Open any project first.

Signing in to your own site

For a site that needs a login, record it signed in. You have two ways. Both save the site's login on the computer that runs the queue.

Saved sign-ins are as sensitive as passwords. Anyone who can read the saved file can act as you on that site. They are plain files, not encrypted. Keep the folder private, sign in only to sites you own or may record, use a test account where you can, and delete a sign-in when you are done (python -m site_video signin https://your.site --forget).

Saved sign-ins live in SITE_VIDEO_SESSIONS (default ~/.site-video/sessions): one JSON file per site, plus a profile folder that holds the sign-in browser's own profile. Delete the whole folder to forget everything.

Option 1: sign in with a browser window (Playwright)

Press Sign in as me on the page, call POST /api/signin, or run:

python -m site_video signin https://your.site

Your installed Chrome or Edge opens on the queue's computer, with a fresh profile of its own. Sign in by hand. Google sign-in works, because it is an ordinary browser. Close the window. site-video then reads that site's cookies and storage and saves them. Your password never passes through site-video. If no Chrome or Edge is found, it uses Playwright's Chromium, where Google sign-in does not work. Set SITE_VIDEO_BROWSER to a browser's program file to pick one.

The window opens on the screen of the computer running the queue. If you are using the queue from another computer, use option 2.

Option 2: the browser extension (your everyday browser)

The extension sends the sign-in your normal browser already has. Nothing closes and no other profile is used.

What it can read, plainly. When you press Send my sign-in on a site, it reads, for that site only:

  • all of the site's cookies, including HttpOnly cookies, which page scripts cannot normally see;

  • the page's localStorage;

  • the page's IndexedDB records that are plain JSON (where Firebase-style "Sign in with Google" keeps its login).

It sends them to the queue address in its Settings, as POST <queue>/api/signin/import, and to nowhere else. It does nothing until you press the button, and only for the tab that is open. It asks the browser for permission to read any site (<all_urls>), cookies, scripting and storage, because it cannot know in advance which site you will press it on. The queue token, if you set one, is kept in the extension's browser storage.

Install it:

  1. Start the queue (python -m site_video serve). Download the extension from the page's "Get the sign-in extension" link, or from http://127.0.0.1:8940/api/signin/extension.zip. If the queue has a token, open the page with ?token= first.

  2. Unzip it. In Chrome or Edge open chrome://extensions (or edge://extensions), turn on Developer mode, press Load unpacked, and pick the unzipped site-video-signin folder.

  3. Open your site in a tab, signed in. Press the extension, then Send my sign-in. It shows how many cookies and stored items were saved.

Where it sends. The download writes the address of the queue it came from into the extension (config.js), so it points at that queue with nothing hardcoded. If that address is not right, change it under Settings in the popup. If you load the folder site_video/browser_extension directly, it starts at http://127.0.0.1:8940.

Plain HTTP warning. If the queue address is http:// and not this computer, the popup shows a warning and asks you to press the button a second time before it sends. Your cookies would cross the network readable by anyone on the path. Use https://, a VPN you trust, or a queue on the same computer.

What the recorder will and will not do

There are two modes. The first is the default.

Default mode: look, never change.

  • The recorder blocks every request that is not a plain page load. Only GET, HEAD and OPTIONS go out. A form submit or any other write is aborted. Trackers and analytics are blocked too.

  • It does not log in by itself. It uses a saved sign-in only if you made one for that site.

  • It moves around the way a viewer would: it scrolls, points at things, opens links and tabs, and types into search boxes. The script is told not to click anything that logs in, buys, deletes or submits a form, and buttons labeled like delete, remove, buy, pay or sign out are never clicked.

  • The limit: a site that changes data on a plain page load (a GET) is not protected by this. Only record sites you may show.

Opt-in mode: your own site, changes allowed. Turn it on with the page's "This is my site" box, --my-site on make and add, or "changes": true in the API. Then:

  • Every request except DELETE goes out, so the video can press buttons that save, create or convert things, as your notes ask.

  • Your notes become the script's workflow: "paste a list, press Evaluate, open the result".

  • Buttons labeled like delete, remove, erase, buy, purchase, pay, subscribe, cancel account or sign out are still never clicked. That is a check on the visible label. It is not a guarantee.

  • Use a test account. This mode really creates and changes things on the site. Pair it with a saved sign-in for the test account, not your real one.

More ways to use site-video

What happens in a video

  1. Explores the site. Every same-site page it can reach by links, up to 60 pages in about 4 minutes (SITE_VIDEO_EXPLORE_PAGES, SITE_VIDEO_EXPLORE_SECONDS). Sign-in, cart and legal pages are left out. Each page is scored for real content, its search fields are tried with an example from the site's own items, and with a vision model set the best views are described.

  2. Reads the brand: background, text and main colors, heading font and logo.

  3. Writes a script with your model: a title card, a "the problem" card, 6 to 10 short scenes that each show one feature working, captions, narration, and a closing card with the site's address. It fits the length you ask for and your notes.

  4. Records the real site in Chromium at 1920x1080, 30 fps, with a soft cursor, eased scrolls and real typing. The waiting is cut out and each scene is sped up (to 2.4x at most) to fit.

  5. Draws the cards and captions in the site's own font and colors, and adds the voiceover.

  6. Edits it in Resolve or with ffmpeg.

  7. Checks the video against the script (needs SITE_VIDEO_VISION_MODEL): each scene is scored 1 to 5 for showing what its caption says. A scene that clearly failed is rewritten and recorded again, once. Verdicts are in review.json.

Notes are a brief. Write a wish ("show pricing") or paste a whole outline with scenes, text to paste and narration. The script follows its order and types or pastes the text you give. Pages named in the notes are always explored.

One video in a terminal

python -m site_video make https://example.com --length 2min --notes "show pricing, end on sign-up" --voice openai:alloy
  • --length: 5min, 90s, 1:30 or auto. Requested lengths are targets: a scene is never shorter than its narration, and the cards add a few seconds.

  • --notes: plain words or auto.

  • --editor: auto (Resolve if it is open, else ffmpeg), resolve or ffmpeg.

  • --model: the model for this video (default SITE_VIDEO_LLM_MODEL).

  • --music file.mp3: music under the voice, quieter while it speaks.

  • --cursor: dot (default), arrow, hand, ring, auto or none. See /cursors on the queue page to compare.

  • --deliver: what to hand back (below).

  • --my-site: the opt-in mode above.

Your own videos

A job can use a website, your own videos, or both.

You have

You give

What happens

A website

the link

The site is recorded

Videos, no website

the videos, each with a note

The script picks moments from your videos, with narration over them or their own sound

Both

both, and a length

Your videos go in where they fit, website scenes fill the rest

python -m site_video make --video demo.mp4 --note-for "the app demo" --video talk.mov --length 2min
python -m site_video make https://example.com --video demo.mp4 --video promo.mp4 --length 5min

Before writing the script it watches each video: a frame every few seconds described by the vision model, and what is said, by time, from speech to text. Without them it knows only your notes and each video's length, so describe the videos well. Your videos are cut, never sped up or slowed. Each one is used at least once.

Changing a finished video

On the page, press Change on a job. Describe the change in words ("in scene 3, point at the Pro plan instead") and press Apply to the script, or edit the script yourself. Then Redo the video: only the scenes that changed are made again. From a terminal: python -m site_video change <id> "point at the Pro plan in scene 3" [--dry-run]. The old script is kept as plan.<n>.json.

Deliverables

--deliver, the page's Deliverables row, or deliverables in the API choose what you get back:

Name

What

video

The rendered MP4 (always)

script

script.md (scenes, captions, actions, narration) and plan.json

voiceover

One WAV per narration line and soundtrack.wav

clips

The recorded scenes and cards, 1920x1080 30 fps

timeline

The Resolve timeline as .drt, .fcpxml and .edl (Resolve edits only)

resolve_project

The video's own Resolve project as .drp (Resolve edits only)

package

One zip with all of the above and a README on importing

From an agent

With the MCP server running and a queue at SITE_VIDEO_SERVER, an agent can call make_site_video, edit_site_video, site_video_status and list_site_videos. The website-promo-video skill teaches it to turn "make a 5 minute video of this site, show X" into a job.

Optional: as an Odysseus mod

odysseus-mod.json and site_video/odysseus_mod.py let an Odysseus server run site-video as a tab and chat tools. You do not need this for anything above.

All site-video settings

Variable

Default

What

SITE_VIDEO_LLM_URL, _LLM_MODEL, _LLM_KEY

local Ollama, qwen3:8b, empty

The script model (see Model setup)

SITE_VIDEO_REPAIR_MODEL

the job's script model

The model that works out a step again when a button is not where the script says (own-site mode)

SITE_VIDEO_DEFAULT_VOICE

none on the command line

The voice when a job names none

SITE_VIDEO_TTS_URL, _TTS_KEY, _TTS_MODEL

OpenAI, empty, tts-1

An OpenAI-compatible voice

SITE_VIDEO_OPENAI_VOICE, SITE_VIDEO_ELEVENLABS_VOICE

alloy, empty

The voice name the page fills in

ELEVENLABS_API_KEY, SITE_VIDEO_ELEVENLABS_MODEL

empty, eleven_multilingual_v2

ElevenLabs

SITE_VIDEO_VISION_MODEL, SITE_VIDEO_REVIEW

empty, 1

Frame descriptions and the scene check

SITE_VIDEO_STT, _STT_URL, _STT_KEY, _STT_MODEL

empty, OpenAI, empty, whisper-1

Speech to text for your videos

SITE_VIDEO_EXPLORE_PAGES, SITE_VIDEO_EXPLORE_SECONDS

60, 240

How far the explorer goes

SITE_VIDEO_HOME

~/site-videos

Jobs and uploads

SITE_VIDEO_TOKEN

unset

If set, the queue's API and page need it

SITE_VIDEO_SERVER

http://127.0.0.1:8940

The queue the command line tools and the MCP server talk to

SITE_VIDEO_MAX_UPLOAD_MB

4096

Largest upload the queue accepts

SITE_VIDEO_MEDIA_DIRS

empty

Folders (; on Windows, : elsewhere) whose videos API jobs may use without uploading

SITE_VIDEO_SESSIONS, SITE_VIDEO_BROWSER

~/.site-video/sessions, auto

Saved sign-ins; the Chrome or Edge used to sign in

SITE_VIDEO_RESOLVE_FOLDER, SITE_VIDEO_START_RESOLVE, SITE_VIDEO_START_RESOLVE_ARGS

Site videos, empty, -nogui

The Resolve edit

FFMPEG

the bundled copy

Your own ffmpeg

The queue's HTTP API is listed at the top of site_video/server.py.

Limits

  • The Resolve edit runs on the computer with Resolve and switches it to the "Site videos" project for the edit, then opens your project again. A Resolve the queue opened is closed afterward; one you opened is left open.

  • Sites that need a login (and you did not sign in), block headless browsers or show a full-screen pop-up the recorder cannot dismiss will not record well.

  • A 30 second request usually comes out at 35 to 40 seconds, because of the cards.

Troubleshooting

You see

What it means and what to do

Could not attach to Resolve

Resolve is not running, or scripting is off, or you have the free edition. Open Resolve, set Preferences > System > General > External scripting using to Local. Free Resolve often cannot do this at all.

Resolve's scripting module is not importable

The server cannot find DaVinciResolveScript. Check Resolve is installed. Set RESOLVE_SCRIPT_API to its Developer/Scripting folder.

No project is open in Resolve

Open a project.

Resolve would not open or create the project folder

Resolve is on its Project Manager screen. Open any project, then try again.

DaVinci Resolve is not running here

You asked for --editor resolve. Open Resolve, set SITE_VIDEO_START_RESOLVE, or use --editor ffmpeg.

No site-video queue at http://...

An MCP tool could not reach the queue. Start it (python -m site_video serve) or set SITE_VIDEO_SERVER.

Executable doesn't exist ... playwright install

Chromium is missing. Run python -m playwright install chromium with the venv's Python (the same one that runs the queue).

Host system is missing dependencies (Linux)

sudo .venv/bin/python -m playwright install-deps chromium

Connection refused or urlopen error while writing the script

The model server is not running or the URL is wrong. Check SITE_VIDEO_LLM_URL with curl <url>/models.

HTTP Error 404 while writing the script

The model name is not on the server. For Ollama run ollama pull <model>, or fix SITE_VIDEO_LLM_MODEL.

HTTP Error 401

The API needs a key. Set SITE_VIDEO_LLM_KEY (or the voice or speech key).

used all N tokens thinking and gave no answer

A thinking model ran out of room. Pick a smaller or non-thinking model.

the model did not write a usable plan

The model could not produce the JSON script in three tries. Use a bigger model. The log shows each attempt.

could not open <url>

The explorer could not load the site: wrong address, no network, or the site blocks headless browsers.

ffmpeg failed: ...

ffmpeg rejected a step. The message ends with its output. Update imageio-ffmpeg or set FFMPEG to a full build with libx264.

token required (401)

The queue has SITE_VIDEO_TOKEN. Open the page with ?token=..., or set SITE_VIDEO_TOKEN for the command line.

a sign-in window for <site> is already open

Close the earlier sign-in window first.

nothing from <site> came with it

The extension found no cookies. Open the site in that tab, signed in, and press it again.

Address already in use

Another program uses the port. Pass --port, or set RESOLVE_MCP_PORT.

PowerShell will not activate the venv

Run .venv\Scripts\python.exe instead of python, or allow scripts for your user with Set-ExecutionPolicy -Scope CurrentUser RemoteSigned.

A job's log is log.txt in its folder, and a failed stage's output is stage.out. A failed job can be retried from its last finished stage.

What is tested, and what is not

  • Tested by the suite: python -m pytest tests runs about 80 tests with a fake Resolve and fake models. They check the MCP server's tools and arguments, the script checker, the queue, the HTTP API, length and voice parsing, the browser extension's address handling, and that no tailnet address or personal path is hardcoded.

  • Run end to end on Linux (Ubuntu 24.04, Python 3.12): the install above, the test suite, the queue and its page, and the MCP server answering resolve_status (it reports connected: false with a reason, as there is no Resolve), and a video made through the ffmpeg editor of a small local page. The language model was a small stand-in server, not a real model: no real model was reachable from that machine.

  • Developed on Windows with Resolve Studio 21.1: the Resolve edit and render. They were not run again while this guide was written.

  • Not tested: start-resolve-mcp.cmd and every Windows command in this guide (the shell version was run); macOS anywhere; Resolve on Linux; free Resolve; the MCP editing tools (add_title, append_clips, codec keys for start_render) against a real Resolve, only against a fake; the browser extension and the sign-in window in a real browser while this guide was written; ElevenLabs, Kokoro and speech to text servers; the Odysseus mod outside its own tests.

License

MIT, see LICENSE. DaVinci Resolve is a trademark of Blackmagic Design. This project is not affiliated with it.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers