Security & Self-Hosting

transcriptcut has no accounts and no login screen. That's a deliberate simplicity choice for a self-hosted, single-local-user tool — not an oversight. It does mean there are things to be careful about once the app leaves your own machine.

No auth in v1 — keep it off shared/public networks

Whoever can reach the app's HTTP port can use it as the one local user: create projects, upload video, trigger renders, read/delete anything. Don't expose the no-auth version of this app to a public or shared network without adding real authentication first. Running it on localhost on your own machine is the intended v1 use case.

The Docker setup follows this: compose.yaml publishes the app on 127.0.0.1:3001, so other machines on your network can't reach it. Changing that to 3001:3001 (or 0.0.0.0) exposes it to everyone on the network, so only do that behind real authentication.

What the app validates, and why it matters more locally

Even without accounts, this is an app that runs FFmpeg and touches the filesystem on your behalf, so it treats client-provided values as untrusted input:

• Project IDs must exist and resolve only to paths under the configured data directory.
• Storage paths are rejected if they'd escape the data directory (no .. traversal) — every file path is validated against the DATA_DIR root before it's read or written.
• FFmpeg parameters are sanitized and validated; user- or AI-provided strings are never interpolated directly into a shell command. FFmpeg always runs with explicit, typed arguments built by application code — see AI-Assisted Editing for the "agents reason, tools execute" rule this follows from.
• Render parameters go through the same validation before a render job is created.

This matters more, not less, for a locally-run tool: FFmpeg is a child process on the same machine as your other data, not a request handled behind a cloud API boundary.

API keys

• .env is gitignored and must never be committed — this is worth double-checking by eye before every commit, not just trusted to .gitignore.
• Your LLM provider API key is used server-side only (in the Bun/Hono backend). It's never sent to the browser.
• .env.example documents every variable without real values, so contributors know what to fill in.

If you're contributing

See CONTRIBUTING.md for the full checklist, and the Community page for where to ask questions. The short version: never commit .env, secrets, data/, the SQLite database file, or real user-uploaded media — test fixtures that do get committed must be synthetic, not real recordings.