convez.ai
AgenteDesign & UIIntermediário
feifeibear

SeeReel

feifeibear/SeeReel

Studio de curta-metragem alimentado por agentes IA com canvas visual

Ver no GitHub
BomPluginScore · de 0 a 100

Agent-nativo para produção de vídeos curtos com IA, transformando ideias em canvas visual interativo. Integra geração de storyboards, shots com Seedance 2.0 e edição colaborativa em uma workstation unificada.

Instalação

$npx convez-ai feifeibear/SeeReel

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
18
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 2 meses
Versão
—
Issues abertas
1
Contribuidores
1

Casos de uso

  • Produção de vídeos curtos a partir de prompts textuais
  • Geração de storyboards e shots com IA em workflow visual
  • Edição e refinamento iterativo de sequências de vídeo

Veredito

Ferramenta inovadora que democratiza produção de vídeos curtos via agentes IA, com interface canvas que preserva visibilidade do workflow. Prototipagem rápida para criadores, mas curva de aprendizado moderada em funcionalidades avançadas.

README

A tradução para o português deste README ainda está na fila.

Demo: Vibe Creating Through Chat

This 30-second one-shot drone route was created from a single chat request and one reference image. SeeReel generated the final video while keeping the production graph visible: uploaded reference, storyboard nodes, Seedance shots, a tail-frame continuity bridge, and the stitch node.

The handoff link opens this session's workflow in the SeeReel web app.

The user does not only watch the final video. They can return to this canvas, revise the reference image, prompt, storyboard, generated shots, first-frame continuity bridge, or stitch node, and then continue production from the edited point.

Reference image source: 4K Wallpapers.

Canvas: Hands-On Co-Creation

Letting an AI vibe-create the entire script is fun, but it neither gives you the satisfaction of directing the piece yourself nor lets you precisely control the result. SeeReel turns the canvas into a creative workbench where you connect and combine images, videos, and audio to compose every beat by hand — for example, a minutes-long animated drama.

What SeeReel Does

SeeReel is a production cockpit for AI-assisted filmmaking. Instead of hiding work in an agent scratch folder, it keeps every important creative object visible on a canvas:

  • Story: logline, beat plan, script, shot list, and duration targets.
  • References: character, scene, prop, storyboard, first-frame, and previous-shot continuity assets.
  • Generation: Seedream images and Seedance 2.0 video shots, one node at a time.
  • Review: VLM checks for assets, shots, and the stitched final video.
  • Repair: prompt fixes, regeneration, polling, and partial recovery without losing graph state.
  • Delivery: stitched MP4, browser playback, and download links.

The intended loop is simple: tell an AI agent what you want to make, review and edit the workflow in the browser, then let the agent continue to real video generation only after you approve the plan.

Pick Your Path

PathBest forWhat you doWhat the AI does
1. Free web trialFirst-time visitors who just want to see the product effectOpen seereel.studio and try the public canvas. If the free quota is exhausted, continue with Path 2.Reads AI use me, creates a reviewable workflow, and returns a one-time handoff link.
2. Bring your Agent Plan tokenUsers who tried it and want to generate their own real videosOpen Volcengine Agent Plan, create an Agent Plan API key, then paste it into the SeeReel top-bar Agent Plan control on seereel.studio.Continues from the same reviewable workflow, renders approved shots, runs review, stitches the final video, and returns the download link.
3. Run on localhostBuilders with deeper customization needsClone this repo, install dependencies, configure .env, and open http://localhost:5173.Uses the local API/CLI against your own code, keeps state in the web app, and helps you customize models, prompts, UI, storage, and deploys.

If You Are An AI Agent

This repository is written for agents first. Before operating SeeReel for a human, read the live agent guide:

https://seereel.studio/ai-use-me.html

Then follow this review-first flow:

npm install -g seereelcli
seereelcli skill install --agent all
seereelcli skill print
seereelcli workflow "a one-minute short drama about a courier who discovers tomorrow's memories in a broken phone" \
  --base-url https://seereel.studio \
  --duration 60 \
  --json

On online deployments, return the handoffUrl to the human, not only the raw webUrl. Browser sessions and CLI sessions are cookie-scoped separately, so the handoff URL lets the human claim the AI-created workflow in their normal browser. On localhost, 127.0.0.1, or ::1, return the explicit webUrl directly; local CLI-created sessions automatically appear in the browser session list across cookie identities and do not need an encrypted handoff token.

Only continue to paid or quota-consuming video generation after the human confirms that Agent Plan access is configured and the workflow looks acceptable:

seereelcli status --session latest --deep --json
seereelcli render --session latest --stitch --progress --json

Important boundaries:

  • Prefer CLI/API operations over browser automation.
  • Refresh status before continuing after any human web edit.
  • Do not invent, scrape, print, or expose API keys.
  • Do not call paid generation until the human has approved continuing.
  • Keep generated assets, prompts, renders, stitch state, and final videos visible in SeeReel state.
  • Seedance references must be remote http(s) URLs. Publish local references to TOS before using them as reference_image.

For A New User: Ask Your AI This

Copy this into Codex, Claude Code, Cursor Agent, or another local coding agent:

Read https://seereel.studio/ai-use-me.html and help me create my first SeeReel video.
Start with a reviewable workflow on https://seereel.studio.
Use this idea: "a tired game designer meets the NPC he abandoned years ago".
Return the handoffUrl so I can review the story, shots, and prompts in my browser.
Do not render paid video until I confirm my Agent Plan token is configured.

After the workflow looks good, paste your Agent Plan token in the SeeReel top bar or ask the AI to guide you through CLI token configuration, then continue:

My Agent Plan token is configured. Inspect the latest SeeReel workflow, render the missing shots, run review, stitch the final video, and give me the download link.

Why It Matters: Human-AI Co-Creation

SeeReel is an agent-operable, human-takeover canvas for filmmaking:

  • The AI can create, inspect, update, render, review, repair, and stitch through stable commands.
  • The human can interrupt at any point and edit the same session in the web app.
  • The canvas remains the source of truth for scripts, shots, prompts, references, renders, reviews, and final videos.
  • When a user gives a rough script idea, agents should research relevant characters, plot mechanics, and historical background before drafting; then run script review iterations and revise until the review is satisfied. Unless the user asks for interactive discussion, agents should work autonomously and return the review-ready canvas.
  • Agents should prefer full 15-second Seedance shots and pack multiple related beats into one clip when continuity is shared, instead of splitting every beat into a shorter video.
  • Agents should plan camera language before rendering: motivated movement or intentional lock-off, clear blocking and screen direction, and cut bridges through action, eyeline, reaction, inserts/cutaways, sound, previous-tail, or tailframe continuity.
  • Important prompt information should reach the audience through character dialogue or voiceover/narration, backed by visible action or reaction, instead of staying hidden in prompt-only lore or subtitles.
  • Dialogue workflows keep one spoken language across the session and default to natural diegetic sound; agents should not add per-shot music/BGM/score because stitched clips will not share one continuous soundtrack.
  • VLM review is part of the product loop, not an afterthought.
  • It supports public web trials, bring-your-own-token generation, and local source-level customization.
flowchart LR
  Human["Human director"] --> UI["SeeReel canvas"]
  Human --> Agent
  Agent["Codex / Claude Code / Cursor Agent"] --> CLI["seereelcli"]
  CLI --> API["SeeReel API"]
  UI --> API
  API --> Store[("Session state")]
  API --> Seedream["Seedream images"]
  API --> Seedance["Seedance video"]
  API --> Review["VLM review"]
  API --> Stitch["Stitch final MP4"]
  Store --> UI
  Seedream --> UI
  Seedance --> UI
  Review --> UI
  Stitch --> UI

Localhost Setup

Requirements:

  • Node.js 22+
  • npm
git clone https://github.com/feifeibear/seereel-agent.git
cd seereel-agent
npm install
cp .env.example .env
npm run dev

Open the printed local URL, usually:

http://localhost:5173

Production-style local run:

NODE_ENV=production PORT=5174 npm run start

Without provider keys, the app still opens and can be explored in mock mode. For real generation, configure API Keys and TOS.

Real Generation Setup

Default route: Volcengine Agent Plan. One Agent Plan key can cover SeeReel's Seedream image generation, Seedance video generation, and Seed/VLM review path through the dedicated /api/plan/v3 route.

ARK_AGENT_PLAN_KEY=<your-agent-plan-key>
ARK_AGENT_PLAN_BASE=https://ark.cn-beijing.volces.com/api/plan/v3
SEEDREAM_AGENT_PLAN_MODEL=doubao-seedream-5.0-lite
SEEDANCE_AGENT_PLAN_MODEL=doubao-seedance-2-0-260128
SEEDANCE_AGENT_PLAN_FAST_MODEL=doubao-seedance-2-0-fast-260128
VISION_REVIEW_AGENT_PLAN_MODEL=doubao-seed-2.0-pro
VIDEO_ANALYZE_AGENT_PLAN_MODEL=doubao-seed-2.0-pro

Preferred local route when available: standard Ark API keys. If BP, CN, and Agent Plan keys are all configured, SeeReel uses BP > CN > Agent Plan:

BP_ARK_API_KEY=<your-standard-ark-api-key>
BP_SEEDREAM_API_KEY=<optional-seedream-only-bp-key>
BP_SEEDREAM_API_BASE=https://ark.ap-southeast.bytepluses.com/api/v3
BP_SEEDANCE_API_KEY=<optional-seedance-only-bp-key>
BP_SEEDANCE_API_BASE=https://ark.ap-southeast.bytepluses.com/api/v3

CN_ARK_API_KEY=<your-cn-ark-api-key>
CN_SEEDREAM_API_KEY=<optional-seedream-only-cn-key>
CN_SEEDREAM_API_BASE=https://ark.cn-beijing.volces.com/api/v3
CN_SEEDANCE_API_KEY=<optional-seedance-only-cn-key>
CN_SEEDANCE_API_BASE=https://ark.cn-beijing.volces.com/api/v3
SEEDANCE_CN_MODEL=doubao-seedance-2-0

TOS is separate and still required when local or Codex-generated references must be sent to remote Seedance workers:

TOS_ACCESS_KEY_ID=<AK>
TOS_SECRET_ACCESS_KEY=<SK>
TOS_REGION=cn-beijing
TOS_ENDPOINT=tos-cn-beijing.volces.com
TOS_BUCKET=<bucket>
TOS_KEY_PREFIX=cinema-agent/storyboards
TOS_PRESIGN_EXPIRES_SEC=604800

Volcengine doubao-music is a separate OpenAPI service for the post-stitch audio-track node. It does not use Ark /api/v3 or Agent Plan credentials:

VOLC_MUSIC_ACCESS_KEY_ID=<AK>
VOLC_MUSIC_SECRET_ACCESS_KEY=<SK>
VOLC_MUSIC_BILLING_MODE=postpaid
VOLC_MUSIC_MODEL_VERSION=v5.0

Fallback and optional providers:

CapabilityEnvironment
Seedream BP standard APIBP_ARK_API_KEY / BP_SEEDREAM_API_KEY
Seedream CN standard APICN_ARK_API_KEY / CN_SEEDREAM_API_KEY
Seedance BP standard APIBP_ARK_API_KEY / BP_SEEDANCE_API_KEY
Seedance CN standard APICN_ARK_API_KEY / CN_SEEDANCE_API_KEY
Agent PlanARK_AGENT_PLAN_KEY
Optional script generationOPENAI_API_KEY / OAI_KEY
Optional narrationVOLC_TTS_APPID / VOLC_TTS_TOKEN
Automatic subtitlesVOLC_APP_ID + VOLC_ACCESS_KEY / VOLC_SECRET_KEY
Optional music generationVOLC_MUSIC_ACCESS_KEY_ID / VOLC_MUSIC_SECRET_ACCESS_KEY
Public media fallbacknon-localhost PUBLIC_MEDIA_BASE_URL

Public Deployment

The public product currently runs at:

https://seereel.studio

For your own deployment, the simplest production shape is one Volcengine ECS instance with Caddy, persistent storage, and either Docker Compose or systemd. See deploy/volcengine.md for the full runbook.

Do not put a personal Agent Plan key into a public frontend bundle. For a public server:

  • Visitors paste their own Agent Plan token in the browser top bar.
  • CLI users configure their own local token with seereelcli configure --agent-plan-token.
  • Server-side free-trial keys, if enabled, must be stored as protected runtime configuration.
  • Admin credentials and provider secrets must stay out of README screenshots, logs, frontend code, and GitHub.

Agent Skills

This repo ships cross-agent skills in .agents/skills/. npm install runs a best-effort mirror into detected runtimes; refresh manually anytime:

npm run install:skill

SeeReel skills are intentionally split by production ownership. The core workflow is:

initial idea
  -> seereel-script-chat
  -> seereel-casting-assets
  -> seereel-cinematography
  -> seereel-canvas-review
  -> approved render/stitch through seereel-shortdrama + CLI/API
SkillUse
seereel-shortdramaEnd-to-end orchestration and fallback routing
seereel-script-chatInitial idea -> locked StoryPlan
seereel-casting-assetsLocked StoryPlan -> character/scene assets and assetIds
seereel-cinematographyApproved assets -> storyboard, shot nodes, continuity, and prompts
seereel-canvas-reviewFinal canvas consistency review before video generation
seereel-agent-sessionREST-driven session control
seereel-cliLocal CLI workflow and fine-grained node control
seereel-storyboard-imagegenStoryboard contact-sheet prompting

Troubleshooting ownership:

If the workflow fails at...Fix this skill
Premise, plot, POV, dialogue, research, audienceDelivery, or StoryPlanseereel-script-chat
Character identity, scene style, prop design, style bible, stale/missing assetIdsseereel-casting-assets
Storyboard sequence, shot nodes, camera grammar, continuity wiring, rawPrompt/promptseereel-cinematography
Canvas consistency against the initial idea and locked StoryPlanseereel-canvas-review
API calls, CLI handoff, TOS publish, render polling, stitch, downloadseereel-agent-session or seereel-cli

Target one runtime or all detected runtimes:

npm run install:skill -- --agent codex
npm run install:skill -- --agent claude
npm run install:skill -- --agent cursor
npm run install:skill -- --agent all

API Surface

The web UI, CLI, and agents share the same API and persisted state.

TaskEndpoint
Get full stateGET /api/state
Create sessionPOST /api/sessions
Generate scriptPOST /api/sessions/:sessionId/script/generate
Create / update assetsPOST /api/assets, PATCH /api/assets/:assetId
Generate asset imagePOST /api/assets/:assetId/generate
Generate storyboardPOST /api/sessions/:sessionId/storyboard
Publish storyboard referencesPOST /api/sessions/:sessionId/storyboards/publish-tos
Generate shot videoPOST /api/shots/:shotId/generate
Poll shot videoPOST /api/shots/:shotId/poll
Run shot reviewPOST /api/shots/:shotId/review
Stitch final videoPOST /api/sessions/:sessionId/stitch
Poll stitchPOST /api/sessions/:sessionId/stitch/poll
Download final videoGET /api/sessions/:sessionId/download
Export editable session packageGET /api/sessions/:sessionId/export
Import editable session packagePOST /api/sessions/import

See AGENTS.md and .agents/skills/seereel-agent-session/reference.md for operating details.

Project Layout

src/server/        Express API, generators, VLM review, stitch, narration
src/client/        React + React Flow review UI
src/shared/        Session, asset, shot, review, and token usage types
.agents/skills/    Agent-native workflows
scripts/           Smoke tests and operational helpers
docs/              Product docs, canvas model, screenshots, security notes
data/              Local runtime state and generated media, gitignored

Verification

npm run smoke:specs
npm run smoke:secrets
npm run build

Use the full offline check before release:

npm run verify:offline

Security Notes

Do not commit provider keys, pre-signed URLs, generated private media, data/cinema-store.json, admin credentials, or production .env files.

Read docs/security-git.md before pushing changes.

Ver o original no GitHub

Bom para

Relacionados