jevreel
MIT MCP and CLI: 1 fps contact sheets, one Jev request per minute to find video moments, verify on full frames, about $0.02 to $0.03 per hour screened.
Overview
jevreel (github.com/DeepAleph/jevreel, MIT, Python 3.11+) from DeepAleph is a video intelligence MCP server and CLI with zero stored embeddings. ffmpeg samples at 1 fps, tiles up to sixty labelled frames per 1568 px contact sheet (one minute per sheet), and sends one Jev request per sheet combining a Noul for whether the moment appears in that minute and a Choice over tile IDs for which second. Top-ranked windows are verified on full-resolution frames every 0.5 s; hits return time, probability, nearby speech, and optional zoom or Box locate tools. README documents ten MCP tools, lens TOML for per-minute typed columns, and five unedited Claude Code sessions on Blender Tears of Steel, including a juice box at 2:28.5 with p 0.95 for 19 requests and $0.0054. README states embeddings win for repeated library-wide search; jevreel targets ad hoc questions at about $0.02 to $0.03 per hour of video screened.
Problem: Agents asked to find a second in a long video either build embedding indexes or burn vision tokens on every frame, with weak probabilities for gating.
Built for: Claude Code, Codex, Cursor, or other MCP clients that need searchable video libraries on disk with timestamps, probabilities, zoom frames, and honest cost tables using a TypeSafe API key.
First indexed on Jev Directory: 2026-10-18
Creator and team
- Name
- DeepAleph
- Handle
- @DeepAleph
How Jev is used
- Role in the product flow
- Parallel per-sheet screening with Noul plus Choice on labelled tiles, then ranked-window verification with per-frame Noul on full frames; optional Box locate and whole-video storyboard Choice
- Primitives
- ChoiceNoulScore
- State in
- Contact sheet images with T01 style labels, transcript or caption text, reference photos for find_in_video, and user question strings; lenses add typed question maps per minute.
- Decision out
- Hits with seconds, probabilities, and frames; timeline rows filterable via query without extra requests; locate returns cropped boxes when edges pass Noul checks.
- uv tool install jevreel[all] with FFmpeg and TYPESAFE_API_KEY
- jevreel index on file or URL builds sheets, transcript, and default timeline questions
- ask_video screens sheets in parallel, verifies top JEVREEL_VERIFY_TOP windows (default 6)
- zoom_video, locate_in_frame, ask_whole_video, and lens tools for follow-up
- MCP server via jevreel mcp or Claude/Codex plugins documented in README
jevreel spends hosted TypeSafe Jev on sheet screening and verification, not on open weights. mcg-nju-videojev trains VideoJev checkpoints that score intervals inside the model; jevreel is an agent tool that keeps vectors out entirely. rayenmlayeh-jevcrawl walks HTML sites with passage scoring; jevreel is for local or URL video files. trungdq88-youtube-sponsor-detection is a Chrome extension that gates sponsor segments in the player; jevreel returns timestamps to your agent for arbitrary questions. jkudish-jev-mcp offers generic rerank and verify tools without ffmpeg indexing. README notes suhailkhalid/framefinder as a community fork only. jevreel is independent of TypeSafe AI; Jev is used through the public API under TypeSafe terms per LICENSE.
Sourced performance claims
- README demo table on Tears of Steel: juice box question found 2:28.5 at p 0.95 with 19 Jev requests and $0.0054; newspaper headline read at 1:20.5 for 8 requests and $0.0024.Source: github.com/DeepAleph/jevreel README See it work
- README cost section measured 2026-10-15 on jev-1.14.0 at $0.042 per million input tokens: one ask screened and verified about 13 to 15 requests and $0.004 to $0.005 on a 10-minute 1080p clip; about $0.02 to $0.03 per hour of video for one question.Source: github.com/DeepAleph/jevreel README What it costs
- README configuration: default JEVREEL_MODEL jev-1.14.0, JEVREEL_FPS 1, JEVREEL_TILES_PER_SHEET 60, verification every 0.5 s in top-ranked windows.Source: github.com/DeepAleph/jevreel README How it works
- Public GitHub repo DeepAleph/jevreel had 31 stars and 3 forks as of October 18, 2026; created 2026-10-15 with CI workflow badge.Source: GitHub API, 2026-10-18
Features and stack
Features
- Ten MCP tools: index_video, ask_video, find_in_video, locate_in_frame, zoom_video, and more
- Lens TOML for noul, choice, and score columns queryable without extra Jev calls
- Agent skill via npx skills add DeepAleph/jevreel; Claude and Codex plugin install paths
- Answer cache so repeated questions on the same sheets cost zero
- Optional yt-dlp URL indexing and faster-whisper ASR with jevreel[all]
- jevreel measure reports per-sheet token cost on your account
Stack
- Python 3.11+
- FFmpeg
- TypeSafe Jev API (jev-1.14.0 default)
- uvx or uv tool install
- Optional faster-whisper and yt-dlp
Pricing: MIT code; Jev input tokens bill per README measured $0.042 per million on jev-1.14.0. Indexing and lens questions can ride in the same sheet request to keep marginal cost low.
Links
FAQ
- Is jevreel affiliated with TypeSafe?
- No. README LICENSE states jevreel is an independent project. It calls TypeSafe Jev through TYPESAFE_API_KEY; it does not ship TypeSafe weights.
- When are embeddings a better fit?
- README Why a decision model instead of embeddings: vector indexes win for querying a large library repeatedly; jevreel pays per minute screened and suits ad hoc questions with explicit probabilities.
- What model version does jevreel pin?
- README configuration table sets JEVREEL_MODEL to jev-1.14.0 on purpose because Box and image tiles need 1.14; aliases may move.
- Can brief events be missed?
- README Limits: verification only walks top-ranked windows (default 6) every 0.5 s; raise JEVREEL_VERIFY_TOP or lower JEVREEL_TILES_PER_SHEET for larger tiles if screening ranks the window low.
- How does jevreel differ from VideoJev?
- VideoJev is an open MCG research model that scores intervals inside the network. jevreel is a Python agent tool that calls hosted Jev on contact sheets and full frames without training checkpoints.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
- Where to run Jev
Compare official TypeSafe, Vercel AI Gateway, OpenRouter, Cloudflare Workers AI, and classifier.dev with a fact table.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.