ReelBrain is a local-first AI video-editing agent team that turns raw footage into creator-review drafts, learns each creator’s taste from explicit feedback, and applies that history to future edits without treating memory as source evidence.
ReelBrain is an OpenAI Build Week submission in the Work and Productivity track.
- Landing page: reelbrain-landing-xi.vercel.app
- Three-minute product walkthrough: Watch on YouTube
- Build Week submission: ReelBrain on Devpost
- Community Edition releases: GitHub Releases
- Four editable editorial personas for story, retention, style, and continuity, coordinated by a Showrunner.
- Local video preflight, playback, fullscreen review, draft version history, and non-destructive revisions.
- Grounded short-form and long-form planning with Korean and English caption artifacts.
- Explicit Yes/No gates for revisions and human approval before generated tools can be built or deployed.
- One-authorized autonomous local-draft mode: the release binds source/runtime/tool scopes, may generate and independently audit only release-supported planner/UI bundles, applies them through compiled host adapters, verifies MP4 output, and stops at creator review.
- Inspectable agent activity, tool calls, orchestration rationale, workflow progress, and durable evidence.
- Like, Dislike, and Skip feedback loops that create explicit preferences, corrections, or tentative taste episodes.
- Creator-controlled Taste Profile records that can be inspected, edited, disabled, or deleted.
- ACP-backed semantic tools and evidence-gated Sleep optimization for prompts, tool descriptions, and workflows.
| Story Editor | Retention Editor | Style Editor | Continuity Editor |
|---|---|---|---|
 |
 |
 |
 |
| Finds a complete educational arc. | Strengthens the hook and pacing. | Applies approved creator taste. | Protects meaning, caveats, and endings. |
The Showrunner receives the creator’s request and selected-draft context, then answers directly or consults the four editors. Approved effects run through bounded semantic tools, produce versioned local artifacts, and remain in creator review until the creator decides otherwise.
The Taste Profile is a behavioral prior, not evidence. A Like becomes an explicit preference with provenance. A Dislike records the creator’s reason and uses it to plan another draft. Skip stores the original request only as a tentative episode; it does not become active taste until consistent evidence exists and the creator confirms it. Current steering always overrides stored taste.
Do not onboard users with the historical GitHub v0.1.1 ZIP or PKG. Those assets predate the bundled-runtime contract and can fall back to an incompatible system Python or missing FFmpeg installation. Distribute only a v0.1.2-or-newer artifact produced by npm run release:macos --prefix desktop, which embeds and verifies Python, Pillow, FFmpeg, and FFprobe before creating the ZIP and PKG. Codex/ChatGPT sign-in or one connected BYOK provider remains the only external model prerequisite.
If macOS says it cannot verify that the unsigned PKG is free of malware:
- Click Done on the warning. Do not move the installer to Trash.
- Open System Settings → Privacy & Security.
- Scroll to Security and find the message saying ReelBrain was blocked.
- Click Open Anyway.
- Authenticate with Touch ID or the Mac password, then confirm Open.
- Complete the normal macOS Installer steps.
Only bypass this warning for an artifact produced from this repository and verified by the release script. The current app bundle is ad-hoc signed, but the PKG is not Developer ID signed or Apple-notarized.
- Choose a video. ReelBrain checks the file locally; no upload or model call happens during preflight.
- Confirm the local engine. A verified release already includes Python, Pillow, FFmpeg, FFprobe, and
uv. If Settings reports one missing, reinstall the verified app instead of asking the user to install developer tools. - Connect one model account. Sign in with Codex/ChatGPT, paste an OpenAI, Anthropic, OpenRouter, or Groq key, or connect a local Ollama engine. Mark exactly one account active.
After those three checks, choose the result, approve the proposed effect, and review the new local draft. Generated-tool provisioning and the one-authorized autonomous path remain Codex-only because they require Codex sandbox and tool-activity semantics; ordinary chat, editorial fan-out, contextual translation, and bounded revision planning use the active provider.
Source-build prerequisites: macOS on Apple Silicon, Node.js, Rust/Tauri prerequisites, and uv. Packaged builds include the local media/Python runtime described above.
git clone https://github.com/Q00/ReelBrain.git
cd ReelBrain
uv sync --dev
uv run reelbrain doctor
cd desktop
npm install
npm run tauri devPrimary creator flow:
- Launch ReelBrain and finish the three checks above. Press ? at any time for keyboard shortcuts.
- Open an existing creator-review project or drag a creator-owned video into Projects.
- Play a draft, use the local timestamp control, and send a revision request to ReelBrain or an
@agent. - ReelBrain checks whether the selected renderer can perform the request before showing the approval card. Unsupported timeline, caption, music, or framing changes stop with a truthful capability explanation.
- Approve a supported request, watch the local render progress, and open the new version.
- Choose Like, Dislike, or Skip, then inspect Memory & History for the resulting provenance.
For the autonomous release path, choose One-authorized autonomous draft after local preflight and Codex connection. The confirmation grants one expiring, budgeted scope for local preparation, four bounded editorial calls, one release-supported generated end-card planner/control bundle, rendering, and verification. ReelBrain may reuse an already active matching tool. It cannot execute an arbitrary generated media capability, and it never publishes; the terminal state is CREATOR_REVIEW.
ReelBrain ships a provider registry, so the editing team is not tied to one vendor:
| Account | Auth | Credential source |
|---|---|---|
| Codex (ChatGPT) | Official Codex browser sign-in | Managed by Codex; ReelBrain never reads ~/.codex/auth.json |
| OpenAI | API key | OPENAI_API_KEY / OPEN_API_KEY |
| Anthropic | API key | ANTHROPIC_API_KEY |
| OpenRouter | API key | OPENROUTER_API_KEY |
| Groq | API key | GROQ_API_KEY |
| Ollama | None (local) | Runs against http://localhost:11434/v1 |
Add accounts in Settings › Accounts. Only the account you mark active receives requests, and you can switch at any time. Keys pasted into ReelBrain go to your macOS keychain; the local registry at .reelbrain/desktop/providers.json stores only an endpoint, a model name, and an opaque keychain reference — never a secret. Credentials resolve in a fixed order: process environment, then keychain, then a project .env. Provider errors are redacted before they reach the interface, logs, or evidence.
Optional provider-backed image and thumbnail features read OPENAI_API_KEY=<YOUR_OPENAI_API_KEY> or OPEN_API_KEY=<YOUR_OPENAI_API_KEY> from a local .env. Do not commit credentials.
- Download the judge sample package from
[SAMPLE_DATA_URL], extract it locally, and keep its directory outside Git. - Or use a creator-owned MP4, MOV, M4V, MKV, or WebM containing both video and audio.
- For the shortest review path, use a 30–60 second rendered draft with adjacent
.en.srt,.ko.srt, and.asscaption files. - For full dogfood preparation, replace placeholders below with local paths and identifiers:
uv run reelbrain dogfood prepare <VIDEO_OR_ZIP_PATH> \
--output <LOCAL_OUTPUT_DIRECTORY> \
--project-id <PROJECT_ID> \
--creator-id <CREATOR_ID> \
--rights-license creator-owned \
--shorts 3 \
--minimum-long-minutes 10 \
--maximum-long-minutes 15Provider confidence does not prove caption accuracy. Correct captions against the source before treating them as creator-approved.
Ouroboros was used to run PM interviews, clarify requirements, generate the ReelBrain Seed, execute the specification, and evaluate the resulting system. Codex with GPT-5.6 then helped turn those decisions into a working Python runtime and Tauri desktop product: it implemented the governed agent and tool architecture, built the dogfood render pipeline, debugged provider and media failures, added persistent taste/evidence flows, and repeatedly exercised the real desktop UI with Computer Use.
The most valuable contribution was not raw code generation. Codex maintained continuity across product decisions, architecture, implementation, tests, and visual QA while preserving the creator’s explicit governance boundaries.
Codex session ID: 019f7e65-0f11-7333-b3fe-c3d8401e0e2a
This session is the primary development record for the Build Week submission. Codex with GPT-5.6 worked across the full product loop rather than only generating isolated code:
- Translated the Ouroboros PM interview and validated Seed into the governed ReelBrain runtime and Tauri desktop application.
- Implemented and debugged four-agent fan-out, Showrunner synthesis, structured human approvals, ACP tool governance, versioned rendering, and creator-controlled Taste Profile evidence.
- Diagnosed real media, transcription, provider, fullscreen, chat, focus, progress, and persistent-state failures while exercising the desktop product through Computer Use.
- Built and ran the Python, TypeScript, and Rust verification suites, then used the results to guide fixes and regression checks.
GPT-5.6 accelerated implementation, debugging, testing, and architectural continuity. The human creator remained responsible for ReelBrain’s product philosophy, governance boundaries, editing-agent model, memory-as-behavioral-prior principle, and final product decisions.
ReelBrain began with an Ouroboros PM interview rather than a prewritten feature list. The interview turned the creator’s initial idea into a validated Seed: solo educational creators, self-contained 30–60 second highlights, accurate bilingual captions, explicit feedback as the basis for taste learning, and the creator as the final source of truth. That Seed became the contract for Codex with GPT-5.6 to implement, evaluate, and repeatedly QA the product.
During development, Ouroboros and Codex fanned out bounded questions to independent subagents instead of asking one agent to reason about every concern at once. Advisory lanes examined code context, challenged assumptions, simplified the workflow, traced architecture implications, and guarded acceptance criteria. The root agent synthesized their evidence, made the implementation change, and ran the relevant tests. This made disagreement visible while keeping one accountable orchestrator.
The same pattern became ReelBrain’s product architecture. A creator request enters the Showrunner, ReelBrain freezes the source catalog, Taste Profile snapshot, workflow epoch, and capability grants, and Codex runs four isolated editorial lanes. Their results are treated as proposals—not authority—until ReelBrain validates their candidate references and the creator approves the effect.
The editable Mermaid source is kept beside the exported diagram.
Codex owns ephemeral execution: spawning lanes, concurrency, retries, and synthesis. ReelBrain owns durable trust: what source was frozen, what each agent was allowed to access, what was attempted, which proposal was accepted, what was rendered, and what the creator chose to remember. Taste is supplied to later agents as a scoped behavioral prior; it never replaces transcript or media evidence.
- Memory is a behavioral prior; it is never evidence for what the source video says.
- The creator is the source of truth for their work, voice, brand, and final approval.
- ReelBrain uses exactly four visible editing personas; the Showrunner orchestrates them rather than acting as a fifth editor.
- Outcomes determine editing quality. Exact tool-sequence assertions are reserved for safety and required evidence steps.
- Natural-language requests are judged by the LLM, not keyword or approval regexes. Effects begin only through structured UI controls.
- Agents may request new tools through ACP, but humans approve quarantined build, testing, and deployment separately.
- Revisions are non-destructive, versioned, locally validated, and never presented as complete until a changed video passes verification.
- Sleep may propose bounded configuration changes, but creator memory, permissions, consent, secrets, and production promotion remain governed.
- The certified development baseline is macOS on Apple Silicon; other platforms are not yet validated.
- Autonomous generated-tool execution is currently release-supported for the end-card adapter; stacked composition can be generated and audited but still requires a second creator-selected video. Arbitrary timeline edits, caption rewrites, music replacement, or new reframing require an additional compiled trusted adapter and cannot become executable merely because code was generated.
- Caption quality still requires creator correction or an independent reference; ReelBrain does not claim automatic 95% accuracy.
- Long-form workflows and provider-backed thumbnail generation require more setup and take longer than the desktop review loop.
- There is no direct social publishing in the current version.
- Taste synchronization, backups, managed rendering, and hosted project history are not part of the local Community Edition today.
The Community Edition is the complete source-available local product: creators can inspect, modify, and self-host it subject to the Sustainable Use License. Local editing, agent configuration, Taste Profile controls, evidence, and governed tools are not intentionally crippled to force an upgrade.
The planned Creator Cloud sells convenience: managed rendering, no API-key setup, persistent Taste Brain synchronization, project history, backups, automatic updates, and creator support. It is not positioned as a gate on the local product’s core editing capabilities.
Issues, reproducible bug reports, evaluation fixtures, documentation fixes, and focused pull requests are welcome. Before contributing:
uv sync --dev
uv run pytest -q
cd desktop
npm install
npm run build
cd src-tauri
cargo testDo not include private footage, transcripts, provider credentials, generated .env files, or unlicensed media in contributions. By contributing, you agree that your contribution may be distributed under the repository’s current license and any applicable contributor terms at [CONTRIBUTOR_TERMS_URL].
ReelBrain is source-available under the Sustainable Use License. Individuals may use, modify, and self-host ReelBrain for personal, educational, research, and internal business purposes, subject to the exact LICENSE terms.
Offering ReelBrain as a competing paid hosted service, reselling it, or commercially providing it to third parties is outside the permitted uses and requires a separate commercial license from jqyu.lee@gmail.com. The LICENSE file is authoritative if this summary and the license differ.