What happened

Shumai covers the usual production workflow: asset storage, frame-by-frame annotations, timestamped comments, share links for clients, and role-based access at team and project level. Storage works on a local filesystem or any S3-compatible backend (AWS S3, Cloudflare R2, MinIO, etc.).

Heavy video transcoding runs in a background worker pool orchestrated by Temporal, so the web UI stays responsive. Custom metadata fields adapt the pipeline without forking the app.

Shumai Agent adds a context-aware chat inside the project workspace, custom tools and scripts, and sandboxed execution for agent-submitted code. Google Gemini can autofill tags and descriptions; vector embeddings power semantic search so you find clips by concept, not just filename.

Three install paths are documented: Docker Compose (download docker-compose.yaml from GitHub), global @shumai-one/shumai via npm/pnpm/Bun, or run from source for development. The npm route needs PostgreSQL with pgvector, ffmpeg, bubblewrap (agent sandbox), and ripgrep. On Ubuntu 24.04+ you may need to relax AppArmor restrictions on unprivileged user namespaces so the sandbox starts.

Why it matters

Frame.io set the UX bar for video review, but it is closed SaaS: data residency, integrations, and scale costs sit with the vendor. Self-hosted alternatives with similar polish are rare; Shumai targets that gap without reducing the product to a bare file drop.

The AI layer is functional: the agent sees project context, runs scripts in a sandbox, and semantic search cuts manual archive hunting. For teams already experimenting with agents in post-production, this is closer to tooling than a demo chatbot.

In practice

1. Quick start — mkdir shumai && cd shumai, fetch the compose file from main, docker compose up -d, open http://localhost:3000.
2. Remote server — set AWS_ENDPOINT_URL_S3 to your public IP or domain in compose so asset URLs are reachable.
3. npm path — run pgvector/pgvector:pg18, set DATABASE_URL and BETTER_AUTH_SECRET in .env, install @shumai-one/shumai globally, run shumai or shumai -d for daemon mode.
4. CLI — upload folders, version assets, and manage projects from the terminal (see the repo CLI readme).
5. Agent safety — verify bubblewrap and sandbox policies on your OS before enabling custom skills.

Takeaway

Shumai is a credible “Frame.io at home”: frame review, permissions, transcoding, and an agent layer in one repo. If you need an open-source creative DAM with collaboration, spin up the Docker demo and compare latency, storage, and data policy against your current SaaS.