Checked, not fully reviewed
A desktop video clipper that can ask Jev, via OpenRouter, whether each clip stands on its own. Checked, not fully reviewed: a very careful Jev client.
Good for
- Long podcasts and streams where an automatic clipper keeps cutting off the setup or the payoff
- Using Jev with only an OpenRouter key, no TypeSafe account
- A reference Jev client: retries, budgets, model checks and a fail-closed policy
Watch out for
- Review & edit always calls Jev; in Automatic it is off by default
- Bounded transcript excerpts and titles go to TypeSafe through OpenRouter
Tested Oct 7, 2026 at b996820ef80f · Python 3.12 in Docker; the engine's Jev, coherence and editorial tests, and its Jev client driven through a mock HTTP transport. The desktop app was not run
Checked, not fully reviewed. We ran the engine’s tests for coherence review, editorial context
and discovery (155, all pass) and drove its Jev client through a mock HTTP transport. We did not run
the desktop app or make any Jev calls.
BridgeClip turns long videos into captioned short clips, on your own machine, with your own
OpenRouter key. Jev is its editor: for each proposed clip it asks whether the cut is faithful to the
source, whether the setup and payoff are both inside it, and whether the title is supported. Each
check has its own minimum probability, and a clip has to pass every check. It reaches Jev as
typesafe/jev-1.13 on OpenRouter’s alpha/decisions endpoint, so no TypeSafe account is needed.
The client is one of the most careful we have read:
| what the mock returned |
result |
an answer from typesafe/jev-1.13-20260918 |
accepted, dated version recorded |
529, then an answer |
answered on the second attempt |
529 three times |
gave up after three attempts: unavailable |
429 asking to retry in 30 s |
gave up at once rather than stall the job (limit 20 s) |
402 |
out_of_credits, and no further requests in the job |
an answer claiming to be typesafe/jev-1.14 |
rejected as an unexpected model |
a state carrying source_context |
the field was removed before sending |
That last one is deliberate: video titles, descriptions and web research are untrusted, so they shape
the clip search but never go to Jev. Requests are capped by count and by a token budget. With a cap
of two, a third distinct question came back budget_exhausted without being sent. And when every
candidate’s review fails, it approves nothing and says why, rather than falling back to unreviewed
clips.
MIT, 364 stars. What leaves your machine is listed plainly in its privacy doc: bounded transcript
excerpts, titles and diagnostic text go to Jev through OpenRouter. Frames go out only if you turn on
the separate visual-context option.
See how it compares with other tools in Best Jev tools, tested hands-on.
Review updated Oct 7, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.