Question 1

Does anything leave my machine?

Accepted Answer

No. The whole pipeline runs locally against a local vision model — no telemetry and no cloud APIs. Your recordings and source code never leave your disk. The HTTP API binds to loopback (127.0.0.1) only. Before anything is persisted, secrets (tokens, keys, passwords) and PII (emails, Luhn-valid cards, SSNs, phones, cloud keys) are scrubbed from every text field, with optional pixel-level redaction on the frames themselves.

Question 2

Does it only work for bugs?

Accepted Answer

No. Framesleuth analyzes any video — a bug recording, a feature demo, a design walkthrough, a Loom, a phone capture — and classifies it (bug / feature / tutorial / demo / feedback / other). The Context Bundle and the action prompt adapt to what the video shows, so a coding agent can fix a bug, add or change a feature, or build a whole new screen or app from the same pipeline. A general (non-bug) video yields a faithful summary plus timestamped key_moments instead of fabricated bug fields.

Question 3

What if I don't have a vision model or ffmpeg installed?

Accepted Answer

Framesleuth degrades gracefully. With only the browser sidecars — console errors, failed requests, clicks — it still produces a valid Context Bundle, and the analysis_quality field honestly reports which stages were degraded. Add a local VLM to unlock frame-level OCR and visual understanding.

Question 4

How do I capture a video in the first place?

Accepted Answer

Any video works — drop in your own file (a screen recording, a Loom, even a phone capture) and the agent analyzes it directly. A browser capture extension can also record a session, collect browser sidecars (console, network, clicks) and post the video to Framesleuth's local API; CORS is already scoped to extension origins and loopback, so that path needs no extra setup.

Question 5

How does it avoid hallucinating what it saw?

Accepted Answer

Every claim in the bundle carries an evidence citation (a frame, a sidecar event, or a transcript span) — uncited claims are dropped. Suspected error frames can be re-read several times and reconciled by majority vote, so one bad read can't decide an error state or inject fake OCR. Agreeing signals across the frame, console and narration reinforce each other via cross-modal corroboration, and each field ships with its own 0-1 confidence plus a task-aware actionability signal (ready / thin / insufficient) you read first.

Question 6

Can I cancel or track a long analysis?

Accepted Answer

Yes. Analysis is async — submit, then poll GET /v1/jobs/{id} or subscribe to a Server-Sent Events stream that closes on completion. Every job reports per-stage timings so you can see where the time went, DELETE /v1/jobs/{id} requests cooperative cancellation at the next stage boundary, a per-job timeout and crash recovery keep the queue clean, and an optional completion webhook plus TTL retention round out the lifecycle.

Question 7

Which models and engines are supported?

Accepted Answer

It's engine-agnostic. The default setup uses Ollama with qwen2.5vl for frame understanding and faster-whisper for narration; swap to llama.cpp or vLLM purely through config — bring the local model you already run.

Question 8

How does it plug into my coding agent?

Accepted Answer

Through MCP. The framesleuth server exposes 14 read-only tools, four resources and a fix_from_video prompt that emits a grounded action prompt (fix a bug, build a feature, or ship something new). Connect it to VS Code Copilot, Claude Code, Claude Desktop or any MCP client; edits only ever happen through that agent's own reviewed apply flow.

Question 9

Can I export an HTML/CSS animation as a video?

Accepted Answer

Yes. The widget's HTML → video mode turns a self-contained HTML animation into a downloadable MP4, WebM, or GIF. With the local agent connected it captures the animation frame-by-frame in headless Chromium (a paused virtual clock advanced one frame at a time) and encodes a color-correct H.264 MP4 — full color, no dropped frames, no quality loss, up to 4K at 5–60 fps. With no backend it falls back to an in-browser preview render. It's the same idea as a dedicated frame exporter, but local and built in.

Question 10

Is it really free to fork and use?

Accepted Answer

Yes — it's open source under Apache-2.0. Clone it, fork it, self-host it, and build on top of it. This marketing site is open source too, so you can reuse it for your own project.

A video is the richestspec you can hand an agent.Framesleuth makes itactionable.

From a raw recording to a fix-ready case file — in nine stages.

Preprocess

Transcribe

Keyframes

Understand

Classify

Extract

Redact

Summarize

Ground

One bundle. The same shape, every time.

Built for real work from video — bugs, features, whole builds — privately, predictably, in your own stack.

Frame-by-frame understanding

Nothing leaves your machine

Engine-agnostic models

Degrades gracefully

Redaction-first

One structured contract

MCP-ready, agent-driven

Grounded in your repo

Auditable by design

Your coding agent drives the whole thing.

Connect your coding agent

Pick how it reads — and what it does next.

Summary skills

Action modes

Your most sensitive recordings never leave your laptop.

Frequently asked

Stop re-typing tickets. Record it once.