videoctx

videoctx

Index a video once. Query it cheaply, forever, with any LLM.

videoctx ("video context") turns any video into an open, portable Video Context index (.vctx, see SPEC.md) using only local, model-free tools: ffmpeg keyframes, PySceneDetect shot boundaries, faster-whisper transcripts, and optional one-line visual captions from any cheap VLM. An agent answers most questions from the index's text layer for fractions of a cent, and pulls individual frames on demand only when a question needs pixels.

videoctx index lecture.mp4          # one-time, local, free

videoctx show lecture.vctx # the text view the model reads videoctx ask lecture.vctx "What diagram is shown when she explains backprop?"

What people use it for

Ask questions about videos the way you'd ask about a document:

• Lectures and courses: "Summarize this 50-minute lecture." "Where does she explain backpropagation, and what's on the slide at that moment?"
• Meetings and interviews: "List the action items with timestamps." "Find the moment the client pushed back on pricing."
• Screen recordings: "At what time does the error appear, and what does the dialog say?"
• Content creation: "Find the three strongest moments for short clips." "Generate chapter markers."
• Video archives: index a folder of talks once, then every future question across all of them costs pennies and answers in seconds.
• AI agents: give any MCP-capable agent (Claude Code, Cursor, ...) the ability to read video at all.

Who this is for: developers and anyone comfortable running a few terminal commands. You'll need Python, ffmpeg, and an Anthropic API key for question-answering (indexing itself is free and offline). If that's not you yet, star/watch the repo: a no-install version is on the roadmap.

Why

Native video ingestion re-tokenizes the whole video on every question: an hour of footage is ~250-350k tokens, every single time. Proprietary video platforms index into closed representations behind per-minute pricing. videoctx is the third option: an open on-disk format any model can consume, produced by free tools, where the marginal cost of a follow-up question on an hour-long video is about $0.02.

Benchmark results (Video-MME, Oct 2026)

Full protocol and honest cost accounting (including prompt-cache write premiums) in paper/draft.md. Query model: Claude Opus 4.8.

Numbers below are single-touch runs of the frozen configs (v0.2: captions + OCR + evidence rules, Opus 4.8 loop; v0.3, 2026-10-21: adds text-evidence rules, verbatim transcript quotes and a final letter-vs-evidence check) under full cost accounting. Where we measured the same configuration more than once, we report the observed range: on n=150, single runs wobble by 2-3 questions; on n=30, by 2-3 questions (= 7-10 points). Treat sub-3-point deltas on this benchmark as noise.

Subset composition caveat: videos were taken in Video-MME dataset order, which clusters by domain. The medium and long cells are entirely Knowledge-domain videos; the short cells are 76-88% Knowledge (remainder Film & Television). Read these as Knowledge-dominated subsets of Video-MME, not the full benchmark; future splits will be domain-stratified.

Short videos, held-out test half (n=150):

Medium videos, 4-15 min (n=30, wide error bars):

Medium is our weakest cell and remains the named target: naive leads on raw accuracy. Ten questions failed identically across five configs (prompt rules, a text-view bug fix, dense captions, a 24-thumb contact sheet); the sixth config, a 60-thumbnail 512px contact sheet baked into the cached prefix, recovered one. The receipts tell a consistent story: these videos are fast-cut and fully captioned, and the remaining gap is visual detail bandwidth at question time, naive's one structural advantage. We publish the ladder rather than only its best rung.

Correction (2026-10-21): a text-view bug had been dropping the header line (duration, shot count) on every OCR-enriched index since OCR shipped; all v0.2/v0.3 numbers were measured under that bug. v0.4 is the corrected configuration (short +2.7 points, long back to 80.0%, medium unchanged).

Long videos, 30-60 minutes (n=30):

Cost correction (2026-10-22): the v0.5 long run started 57 minutes after the v0.4 run, and for the 7 of 10 videos whose text views were unchanged by dense captioning, it silently inherited v0.4's still-live 1-hour prompt caches, flattering first-question costs. $0.109/q is the cold-equivalent (first question $0.285, follow-ups $0.021 as measured). Benchmark runs are now cache-salted per config so this cannot recur. | Uniform frame sampling (60-frame cap) | 70.0% | $0.133 | $0.133 | | Gemini Flash-Lite native | 76.7% | $0.022 | $0.022 (5-9 min latency/q) | | Native ingestion (published Pro rates, est.) | pending | ~$0.50 | ~$0.50 |

One-time index cost: ~$0.10-0.25 per hour of video (captions + local OCR + local embeddings), pennies otherwise.

*Cost correction history: early short-video figures ($0.0067/q) undercounted prompt-cache write premiums; the metering was fixed 2026-10-18 and all figures above are fully loaded. Details in the paper's Section 7.3.*

Install

git clone <repo> && cd videoctx

uv venv && uv pip install -e .

Requires ffmpeg on PATH. videoctx index and the MCP server's core tools run fully locally with no API key. videoctx ask uses the Anthropic API (ANTHROPIC_API_KEY or an ant auth login profile).

MCP server: let any agent read videos

claude mcp add videoctx -- $(pwd)/.venv/bin/videoctx-mcp

The calling agent's own model reads the text layer and frames directly; the server makes no model calls and needs no key. Tools: video_overview (cheap text layer), survey_video (one thumbnail per shot), get_frames(path, t0, t1, detail), transcript_search, find_moments (semantic visual search, needs videoctx embed once per index), index_video. The server's instructions carry the evidence rule: visual answers must cite fetched frames, never general knowledge.

Reproduce the benchmark

uv pip install -e ".[eval]"

python eval/prepare_videomme.py --duration short --videos 100 --out eval/data python eval/split_dev_test.py eval/data/videomme_short.jsonl python eval/run.py eval/data/videomme_short_test.jsonl --methods videoctx,naive

eval/methods.py documents each baseline. Cost accounting counts cache writes at their real premium (1.25x / 2x by TTL); see the paper's Section 7.3 for the accounting traps this avoids.

Format

The .vctx directory is plain JSON + JPEG: manifest.json (provenance, parameters, source hash), index.json (shots, keyframes, transcript, captions), frames/. Consumers must ignore unknown keys; enrichment layers (OCR, embeddings) are reserved in SPEC.md v0.1.

Status and roadmap

Working and benchmarked: indexing, query loop, MCP server, eval harness. Next: medium-duration subset, Pro-tier native baseline, cross-model portability runs (same indexes via non-Claude models), OCR and per-shot embedding layers, spec v1.0.

License

Apache-2.0. Copyright 2026 Fred Dunyo.