Video Agent Runtime

Agent-native video editing, built around durable plans, reviewable versions, and deterministic rendering.

Edit video as structured data — not as opaque shell commands.

Quick start · Agent & MCP · Speech & voice · Mobile · Documentation · Benchmarks

What is it?

Video Agent Runtime is a headless editing engine for agents such as Claude Code, Codex, or any MCP-capable client.

Instead of letting a model directly author FFmpeg commands, it turns media into durable project data:

Transcript → EditingStrategy → EditPlan / EditPatch → Timeline → Version → Preview → Approval → Export

The model decides what should change. The runtime validates whether that change is allowed, applies it transactionally, renders the result, and keeps every mutation reviewable and reversible.

It is designed first for talking-head video, interviews, podcasts, lectures, screen recordings, and long-form-to-short-form workflows.

Related MCP server: video-editor

How it works

flowchart LR
    A[Source video] --> B[ASR + visual evidence]
    B --> C[Transcript / Timeline context]
    C --> D[LLM proposes EditingStrategy]
    D --> E{User approves?}
    E -- no --> D
    E -- yes --> F[Structured EditPlan]
    F --> G[Validate + Diff + Apply]
    G --> H[Immutable Version]
    H --> I[Preview render]
    I --> J{Review}
    J -- feedback --> K[EditPatch or Replan]
    K --> G
    J -- approve --> L[Final export]

A typical interaction looks like:

Import interview.mp4
↓
"剪成一分钟，开头抓人，删掉废话"
↓
Agent proposes a hook-first strategy
↓
User approves
↓
Runtime validates and applies an EditPlan
↓
Preview
↓
"前 20 秒还是太慢"
↓
Minimal EditPatch → new Version → new preview
↓
Final approval → export

Core capabilities

Speech & voice

Speech is a first-class editing subsystem, not a subtitle add-on. ASR produces the semantic timeline used by the editor; TTS and VoiceProfile outputs become explicit project assets and timeline clips.

ASR

TTS and voice identity

Voice cloning is never automatic. A cloned VoiceProfile requires explicit authorization evidence, a quality-checked reference, and—where the provider supports it—an exact transcript-backed reference range. Multi-speaker media is not silently guessed.

Model code, weights, voice assets, and hosted APIs may have different licenses or commercial terms. See speech model research and voice identity before shipping a provider configuration.

Agent & MCP

All public surfaces are thin adapters over the same VideoAgentCore; they do not maintain separate project or timeline models.

Connect Claude Code / Codex

Build the repository, then point an MCP-capable client at the stdio server:

npm install
npm run build

mcp.example.json contains a minimal configuration shape. Provider secrets are read from environment or secure host storage and are never written into project JSON.

Quick start

Requirements

• Node.js 22+

• FFmpeg / FFprobe for real media rendering

• Optional Python environment for local speech models

Install and verify

npm install
npm run typecheck
npm test
npm run build
npm run smoke:mcp
npm run demo

npm run demo creates synthetic source media locally and drives the real project workflow through transcript → strategy → versions → previews → feedback patch → narration → final FFmpeg export.

Check the current machine without making a paid model call:

npm run cli -- doctor

Configure providers

Copy the relevant values from .env.example into your environment.

# Workspace
VIDEO_AGENT_WORKSPACE=./video-projects

# Planner
VIDEO_AGENT_PLANNER=openai
OPENAI_MODEL=gpt-5.4-mini
OPENAI_API_KEY=...

# Local ASR example
VIDEO_AGENT_ASR=qwen3-asr
VIDEO_AGENT_ASR_MODEL=Qwen/Qwen3-ASR-0.6B

# Local TTS / voice example
VIDEO_AGENT_TTS=qwen3-tts
VIDEO_AGENT_TTS_MODEL=Qwen/Qwen3-TTS-12Hz-0.6B-Base
VIDEO_AGENT_PYTHON=python

Other supported choices are documented directly in .env.example.

Real-provider validation

Normal CI intentionally does not download large speech models or use paid credentials. Real providers are verified through an explicit acceptance harness:

VIDEO_AGENT_REAL_ACCEPTANCE=true \
VIDEO_AGENT_ASR=qwen3-asr \
VIDEO_AGENT_PLANNER=openai \
VIDEO_AGENT_TTS=qwen3-tts \
OPENAI_API_KEY=... \
npm run eval:speech-real

The harness records real stage latency, ASR/TTS real-time factor, provider/model metadata, Node-controller RSS and coarse GPU memory when available. Authorized voice-clone acceptance must be enabled separately and cannot silently run against an arbitrary speaker.

After several runs, aggregate comparable results with:

npm run benchmark:speech-summary

See real speech acceptance, benchmarks, and speech model research.

Mobile host

The mobile target is designed around a local-first, zero-application-server architecture:

Mobile App
  ├── VideoAgentCore
  ├── durable ProjectRepository
  ├── Workflow / Job Queue
  ├── Timeline / EditPatch / Version
  ├── native media adapters
  └── direct BYOK provider access when configured

Source media stays on device by default; remote providers receive only the approved ContextPack/evidence required for inference. API credentials are referenced through secure host storage rather than project JSON.

Start at docs/mobile/README.md and native host status.

Architecture

                 Agent / CLI / MCP / Mobile
                          │
                          ▼
                    VideoAgentCore
                          │
        ┌─────────────────┼──────────────────┐
        ▼                 ▼                  ▼
     Workflow          ProjectStore       Job Queue
        │                 │                  │
        └──────────┬──────┴──────────┬──────┘
                   ▼                 ▼
              Edit / Timeline    Provider contracts
                   │                 │
             Version / Diff     ASR / LLM / TTS
                   │                 │
                   └────────┬────────┘
                            ▼
                         Renderer
                            │
                         Preview
                            │
                      Review / Export

The main invariant is simple: the runtime owns state; models propose structured changes.

For package boundaries, persistence layout and recovery semantics, read architecture.md.

Documentation & development

The README is the product entry point. Technical details live under docs/.

Project status

The Node runtime is the verified primary path: CLI, MCP, durable project state, FFmpeg rendering, jobs, deterministic evaluation, speech provider adapters and real-provider acceptance tooling are implemented and covered by CI where they do not require external model weights or paid credentials.

Real local-model quality, latency, VRAM and hosted-model behavior must still be measured on the target machine through the opt-in acceptance harness; CI does not pretend those runs happened.

The mobile host remains a source-level prototype until native compilation and real-device validation are completed.

Security principles

• Agents never receive arbitrary shell execution through the editing API.

• Raw FFmpeg strings are not authoritative edit state.

• API keys are never persisted in project JSON, ProviderCall records or benchmark reports.

• Source media remains local unless a workflow explicitly authorizes remote evidence.

• Voice cloning requires explicit authorization and provenance.

• Unsupported renderer/provider capabilities fail explicitly rather than degrading silently.

See docs/security.md for the full boundary.

License

MIT. See LICENSE.