DaVinci AutoEdit Agent: a Codex skills repo that builds Resolve timelines from raw footage

The gap DaVinci AutoEdit Agent fills: footage in, approved timeline out

Most editing automation stops at the interesting part. A tool will transcribe, or cut silence, or generate a rough assembly, and then hand you a file you still have to rebuild by hand inside your NLE. DaVinci AutoEdit Agent aims at the whole chain instead. The README describes a pipeline that starts with arbitrary video, audio and image media and ends with a Resolve timeline, a delivery audit and a pickup-shot report. The repository explicitly makes no assumptions about media location, camera, subject, language, platform or duration.

The intended user is someone who already edits in DaVinci Resolve and already runs Codex with local skill support. It is not aimed at people who want a hosted web editor. The project ships as three skills rather than one monolith: davinci-autoedit-agent for the end-to-end run and its approval gates, viral-video-writer for narration, hooks and titles grounded in actual material evidence, and davinci-resolve-editor for operating Resolve through MCP and verifying that writes landed. That split matters because you can invoke the writer or the Resolve operator on their own without starting a full production pass.

How the pipeline actually runs, stage by stage

The workflow is a linear sequence of named stages, and the README lists them in order. The agent explains the process, confirms media paths, subject and delivery goal, asks about your pacing and editing practices, checks LLM, TTS and Resolve configuration, scans video, audio and images, extracts frames for analysis and tagging, then asks you to confirm the material review. Only after that does it draft and confirm a script, optionally configure and generate TTS, draft and confirm an edit blueprint, and create the Resolve project and timeline. A final audit checks gaps, adjacent repetition, source bounds and write results, then compares the finished cut against the script to produce a pickup-shot list.

The architectural detail worth noticing is where the intelligence sits. Frame extraction and scanning rely on FFmpeg and FFprobe on PATH. Script and blueprint generation can use an OpenAI-compatible multimodal endpoint, but the README states that without an LLM API you can fall back to Codex's own vision capability or a manual review worksheet. TTS is likewise optional and points at a user-supplied HTTP API. Resolve access goes through the upstream davinci-resolve-mcp server, which the README tells you to install separately so it keeps receiving its own updates. There is also a documented Resolve Python fallback for when MCP is unavailable, used only after user confirmation.

The stated principles read like a list of failure modes the author has already hit. Originals are immutable. Every creative stage needs approval before it is written. A source family is never rejected wholesale because early samples scored badly. The audit covers coverage, adjacent repetition, source bounds, frame origin, gaps and final project identity. And there is an explicit rule against reporting a successful measurement as a successful write, which is the classic bug in any tool that talks to an application through a scripting bridge. Structural revisions go into a new project or timeline, and user-owned audio, grades, BGM and tracks are preserved unless authorized.

Installing the skills and running a first edit

The README gives a four-command install. Cloning, installing requirements and running the installer script register the skills so Codex can see them. Note the directory name in the second command differs in case from the repository name, which is how the README writes it.

 git clone https://github.com/liuluhaixiu/DaVinci-AutoEdit-Agent.git
 cd davinci-autoedit-agent
 python -m pip install -r requirements.txt
 python scripts/install_skills.py

If you would rather not run the installer by hand, the README offers a Codex prompt that points at the repository URL and asks for all skills to be installed. Either way, Resolve connectivity is a separate step, and the README is specific that you install the upstream MCP server yourself so it continues to receive updates.

npx davinci-resolve-mcp setup

API-backed features are optional, so configuration is opt-in. Create the environment file only when you need an LLM or TTS endpoint, and fill in the keys you actually have.

cp .env.example .env

The example file defines LLM_BASE_URL, LLM_API_KEY, LLM_MODEL, LLM_SUPPORTS_IMAGES and LLM_TIMEOUT_SECONDS for the model side, and TTS_BASE_URL, TTS_API_KEY, TTS_MODEL, TTS_REFERENCE_AUDIO and TTS_TIMEOUT_SECONDS for narration. A separate pair, RESOLVE_SCRIPT_API and RESOLVE_SCRIPT_LIB, covers the Resolve fallback. Windows users are given a PowerShell equivalent, Copy-Item .env.example .env.

Before starting a real job, run the readiness check. It is the fastest way to find out that FFprobe is not on PATH or that your Resolve build will not accept external scripting.

python skills/davinci-autoedit-agent/scripts/check_setup.py

Then invoke the skill with a plain-language brief. The README's own example names a footage folder, a music folder, a duration, a platform, a subject, and two editorial choices: restrained pacing and J-cuts, with narration switched off.

Use $davinci-autoedit-agent.
My footage is in D:\Media\launch-event and my music is in D:\Music.
Make an 8-minute YouTube documentary about the people behind the event.
Use restrained pacing, J-cuts, and no narration.

Expect the first response to be a confirmation exchange rather than an edit. The agent asks about your pacing and editing practice once; if you have none, it falls back to its documented best-practice profile. From there it moves through material scanning and frame analysis, and it will pause at the material review, the script and the blueprint for your approval.

Where the approval gate becomes a cost

The same design decision that makes this safe makes it slow. Every creative artifact is previewed in chat and needs explicit approval before it is written or used by the next stage. On an eight-minute documentary that is at least three review points, and each one is a conversation rather than a button. If your footage is small and your intent is obvious, the gate is overhead you did not ask for.

The harder limit is Resolve itself. The README states that the free edition of Resolve may not expose external scripting. When that happens, the agent can still do material analysis, writing and blueprint generation, but it cannot build the timeline for you. You get a plan, not a cut. Anyone evaluating this as a free alternative to manual editing should read that line twice.

There is also a dependency shape worth flagging. The project pins almost nothing: requirements.txt lists python-dotenv and requests, and everything else is either a system binary you supply or an upstream server you install separately. That keeps the repository small, but it means a broken FFmpeg install or a stale MCP server shows up as an agent failure rather than an obvious package error. The check_setup.py script exists precisely because that diagnosis is otherwise unpleasant. Finally, the README does not document rollback behaviour for a partially written timeline, and no releases are listed in the repository, so there is no versioned changelog to consult before upgrading.

How it differs from Resolve's own scripting and from MCP-only setups

The closest alternative is not another AI editor. It is driving Resolve directly, either through the Resolve Python API or through davinci-resolve-mcp on its own. The difference is where the decisions live. With a raw MCP setup, you or your own agent decide what to cut and when; the server just exposes Resolve's capabilities as callable tools. DaVinci AutoEdit Agent sits above that and supplies the editorial sequence: scan, tag, review, script, TTS, blueprint, build, audit. It also supplies the guardrails, including the rule that originals stay read-only and the requirement that structural changes go into a new project or timeline.

That layering is the honest selling point and the honest limitation. If you already have a working agent that talks to Resolve and you trust your own prompts, this repository adds a fixed workflow you may not want. If you have never wired an agent to Resolve, the bundled davinci-resolve-editor skill and the write-verification rule are the parts that save you from silently corrupting a project. The project also deliberately uses the upstream MCP server rather than vendoring a fork, which means you inherit upstream fixes but also upstream breakage.

Licence, maintenance and what an upgrade actually involves

The repository is MIT licensed, which permits commercial use and modification provided the copyright notice and permission notice are retained. It also ships a THIRD_PARTY_NOTICES.md, and since the project depends on FFmpeg, the upstream davinci-resolve-mcp server and any LLM or TTS endpoints you configure, those components carry their own terms. That is a compliance question for your own counsel, not something the README resolves.

On maintenance: the repository is not archived, and the last push was on 2026-06-12. No releases were retrieved, so there is no published version to pin against. Practically, an upgrade means pulling the branch and re-running python scripts/install_skills.py, because the skills are registered from the working tree rather than distributed as a package. Because requirements.txt pins only python-dotenv and requests, upgrading will not surface a dependency conflict for you; the moving parts are FFmpeg, your model endpoint and the separately installed MCP server. Budget for re-running check_setup.py after any of those change.

Editorial conclusion

Adopt it if you already run Codex with local skill support and want a staged, auditable path from a folder of footage to a Resolve timeline, and if you can live with approving each stage in chat. Do not adopt it if you need a one-click renderer or you work on Resolve Free, since the README states the free edition may not expose external scripting. Verify three things before committing: that python skills/davinci-autoedit-agent/scripts/check_setup.py reports your FFmpeg, LLM and Resolve configuration as ready, that your Resolve build is Studio 18.5 or newer, and that npx davinci-resolve-mcp setup completes against your installed version.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. liuluhaixiu/DaVinci-AutoEdit-Agent on GitHub
4. README