jevreel

A zero-embedding approach to video intelligence, using Jev.

jevreel lays a video out on labelled contact sheets, one frame per second and sixty frames to an image, and has Jev, TypeSafe's fast decision model, judge every sheet in parallel. The best windows are re-checked on full frames. Your agent (Claude Code, Codex, Cursor, Claude Desktop, anything that speaks MCP) gets back moments, times and probabilities, then looks closer itself. No vectors are computed or stored anywhere.

See it work

Five unedited Claude Code sessions using the jevreel MCP server and skill on Blender's open movie Tears of Steel (12:14, (CC) Blender Foundation, CC BY 3.0). Each clip is drawn from its recording: the agent's tool calls, what Jev answered for every sheet and frame, the hits, the cost, and the agent's own answer, quoted. Click a clip for the full-resolution video.

1. Find a three-second moment

2. Search the dialogue

3. Read what's on screen

4. Find a person from a photo

The reference photo is a crop of the frame at 9:18, so the 9:17 hit is its own source; the other seven are what the search found.

5. Map a whole film

The session starts with two failed calls, left in: a shell command the recording didn't allow, then a wrong folder. The error names the folder relative paths resolve from, and the next call works.

Why a decision model instead of embeddings

Multimodal embedding models already search video well: EmbeddingGemma 2 (October 2026) puts text, images, video frames and audio in one vector space and runs on a phone. jevreel takes a different route to the same goal, with different trade-offs.

• Your question, not a fixed notion of "similar". An embedding index decides at index time what counts as close. A decision model judges the question you ask now, and returns a probability your code can threshold, sort and gate on.
• Every second, cheaply. Jev charges for input tokens only, and output is free. A contact sheet is sent once and every question about it rides along in the same request, so jevreel screens the whole video at one frame per second instead of a handful of keyframes.
• Questions are nearly free, so ask many. Lenses add your own typed questions to every minute at almost no extra cost, then filter them as often as you like for nothing.
• The expensive model only sees what passed. Your agent gets the few moments that matter, as full-resolution frames, rather than a pile of thumbnails.
• Typed answers, not just rankings. Yes/no with the second it happens, one of your options, or a level on your scale, per minute or for the whole video.

Embeddings are the better fit for querying a large library over and over: once the vectors exist, each search is a lookup that costs nothing, while every jevreel question pays Jev for the minutes it screens (about $0.02–0.03 per hour of video). The two combine well: embeddings to shortlist videos, jevreel to find, verify and describe the moment.

How it works

flowchart LR
    V[video file or URL] --> F["ffmpeg: 1 frame/s"]
    F --> S["contact sheets<br/>60 labelled tiles, 1 minute each"]
    V --> T["captions, or local<br/>speech-to-text"]
    S --> J{{"screen: one Jev request<br/>per sheet, in parallel"}}
    T --> J
    J --> R["verify: top-ranked windows<br/>re-checked on full frames,<br/>every 0.5 s around the pick"]
    R --> H["hits: time, probability,<br/>frame, nearby speech"]
    H --> A["your agent: zoom, locate,<br/>read, cite 12:04"]

1. Index. Sample the video at 1 fps and tile the frames onto 1568 px contact sheets, each tile labelled T01, T02, and so on. Fetch captions, or transcribe locally. Ask every sheet a fixed set of questions once (people, on-screen text, cuts, setting, activity, a representative frame) plus any lenses you add, which gives a typed timeline.
2. Screen. One request per sheet: a yes/no (Noul) for "is it in this minute" plus a Choice over tile IDs for "which second". Screen and speech are separate questions combined in code.
3. Verify. The top-ranked windows (by rank, not threshold) are re-checked with one yes/no per full frame, every half second around the pick. Screening ranks a one-second event correctly even when it can't confirm it; verification confirms it, or throws out a false lead.
4. Look closer. zoom re-renders full-resolution frames; locate puts a Box around a target in one frame and crops it; whole answers one question about the entire video from a storyboard, the timeline and the transcript.

Answers are cached: the same images and questions to the same model never cost twice.

This follows TypeSafe's own patterns. Their line-search cookbook pairs a Choice over labelled units with a Noul for "is it there at all". Their images guide searches a large image by Choice over tiles and looks closer only where needed. Their Box guide gates every box behind a Noul and checks its edges.

Quickstart

You need Python 3.11+, FFmpeg on your PATH, and a TypeSafe API key.

uv tool install "jevreel[all]"
export TYPESAFE_API_KEY=...

jevreel index talk.mp4                       # prints the video id
jevreel ask talk "a slide with a bar chart"  # id, id prefix, path or title all work
jevreel zoom talk 12:00 12:10                # full-resolution frames to look at

[all] adds URL support (yt-dlp) and local speech-to-text (faster-whisper). Without them, jevreel indexes local files and uses .srt or .vtt captions saved next to the video.

Lenses: your own questions on every minute

A lens is a set of typed questions asked of every sheet and kept as timeline columns. Questions can be noul (yes/no, plus the second it happens), choice (one of your options) or score (levels, lowest first).

# examples/lenses/driving.toml
name = "driving"

[[questions]]
name = "on_bridge"
type = "noul"
ask = "Is the car driving on a bridge?"

[[questions]]
name = "weather"
type = "choice"
ask = "What is the weather mostly like?"
options = ["clear", "cloudy", "rain", "snow", "fog"]

[[questions]]
name = "traffic"
type = "score"
ask = "How much other traffic is on the road?"
levels = ["None", "Light", "Moderate", "Heavy"]

jevreel index drive.mp4 --lens driving.toml     # rides in the indexing requests: almost free
jevreel lens drive driving.toml                 # or run it later: one request per minute of video
jevreel query drive "on_bridge > 0.7"           # filter as often as you like: no requests
jevreel query drive "weather == snow and traffic >= 2"

Use it from your agent

The MCP server exposes ten tools. The skill teaches agents the workflow: ask, look before claiming anything, cite times.

Claude Code

claude mcp add jevreel -e TYPESAFE_API_KEY=your-key -- uvx --from "jevreel[all]" jevreel mcp

Codex

codex mcp add jevreel --env TYPESAFE_API_KEY=your-key -- uvx --from "jevreel[all]" jevreel mcp

Claude Desktop, Cursor and other mcpServers clients

{
  "mcpServers": {
    "jevreel": {
      "command": "uvx",
      "args": ["--from", "jevreel[all]", "jevreel", "mcp"],
      "env": { "TYPESAFE_API_KEY": "your-key" }
    }
  }
}

VS Code uses the same entry under "servers" with "type": "stdio".

The skill (Claude Code, Codex and other agents that read Agent Skills):

npx skills add DeepAleph/jevreel

Agents without MCP can use the skill with the CLI alone; every command prints JSON.

Tools

jevreel measure <video> reports what one contact sheet costs at several grid sizes on your account.

Python

import asyncio

from jevreel import Reel


async def main() -> None:
    async with Reel() as reel:
        video = await reel.index("talk.mp4")
        result = await reel.ask(video.id, "a slide with a bar chart")
        for hit in result.hits:
            print(hit.time, hit.probability, hit.frame)
        print(reel.query(video.id, "text_on_screen > 0.8").rows)


asyncio.run(main())

Configuration

What it costs

Measured on 2026-10-15 with jev-1.14.0 at $0.042 per million input tokens, on a 10-minute 1080p video:

One question across an hour of video comes to about $0.02–0.03. Asked at index time, lens questions ride along with the default ones, which is where the "nearly free" comes from: an 8-question request cost about the same as a 2-question one. TypeSafe doesn't document image token counts yet. These figures are what the API reported. Check yours with jevreel measure.

Measured on a real video

A 10-minute 1080p gameplay capture, no speech, every result checked against the frames by hand:

Limits

• Brief events need screening to rank their window. Verification looks every half second, but only in the top-ranked windows (6 by default). An event that screening ranks below that is not checked. Raise JEVREEL_VERIFY_TOP, or lower JEVREEL_TILES_PER_SHEET for larger tiles.
• No sound beyond speech. Audio reaches Jev only as transcript text.
• Jev's image support is experimental. TypeSafe calls images and Box "dreamer-ready", not production-ready, and says limits and accuracy can change between releases.
• Jev's real image minimum is 280 px per side. The docs say 256, but the API rejects anything smaller than 280 (measured 2026-10-15). jevreel pads every image itself.
• Thresholds need tuning on your footage. 0.5 to report and 0.7 for "strong" follow TypeSafe's examples. They held up on the test video, which is one video.

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md. Coding agents should read AGENTS.md.

License

MIT. jevreel is an independent project, not affiliated with TypeSafe AI. Jev is TypeSafe's model and is used through its public API under TypeSafe's terms.