Introducing jevgrep - a research agent CLI powered by jev from TypeSafe AI that reduces your coding agent cost by 40% (verified on SWE-bench) Make sure to use the built in skill so your coding agent knows to use jg for context collection

This works because coding agents typically spend 30-60% of all its tokens on research to collect context before writing a single line of code. The actual code generation tokens are tiny. To install, just send claude/codex this exact repo (or even the exact tweet)

A lot of time was spent on optimization, gpt-6-astra ran this in an autoresearch loop to optimize cost / perf for ~70hrs to find the optimal input & output shapes for API calls & the CLI This CLI is designed 100% for agents, its outputs would make no sense to a human

Also - this video (including the sound) is made 100% in opus-5.5, what a time to be alive!

On why this implementation is different compared to other general Jev based file search tools:

@typesafeai related :)

Took a quick look - this is a good generlized Jev implementation, but it won't work for coding agents, you need a recusive code discovery loop (e.g. an actual research agent) to make the context useful, else it'll be either too much context or too little and won't be useful enough to cut costs This would be good for humans where my jevgrep is made 100% for agents. You can try to use the cli yourself but the output will be too dense & confusing.

@eltokh7 @typesafeai Yes I found exactly this - for reviewing code, even if you pass a ranked list of hunks in as context, the agent is still gonna just pull a huge chunk of the file context in anyway. Fighting the weights

@eltokh7 @typesafeai Yup which is why the cli output needs to contain instructions for the agent and structured to be agent friendly, and also why the built in skill is important

@typesafeai Does it pass the @theo test?

@typesafeai @theo What test is that 😅

i'd love to test it out if you want to truly take it a step further find a local classifier model (there are alot emerging every day) laya is the 1st that i can think of and let astra / opus 5.5 post train it (if it's doable) that's an experiment i might run myself if i found jevgrip useful haha

@typesafeai Yea will def be testing out diff models

@typesafeai Great man! I will do a Eikosgrep forking it!

@typesafeai 40% is impressive, but isn't swe-bench retard? need the skill to generate this motion video ASAP!

@typesafeai deepswe is better but that's like $500 per run. Swebench is the poor man's benchmark 😂

@typesafeai very cool!

@typesafeai Thanks! Give it a try, it's been making my max plans last a lot longer

@typesafeai Cool, I want to try this

@typesafeai Lmk how it works out for you!

@typesafeai a 40% cost reduction is impressive, tools that optimize coding efficiency are essential

@typesafeai Simple but strong idea

@typesafeai swe bench is trash

@typesafeai I know 😅

@typesafeai LOVE this 💪 would be really cool to see a fuller eval comparing harnesses using ripgrep vs jevgrep

@typesafeai If there's enough interest I will def put in some more $$ for a full run & with other models

@typesafeai How to make a promo video like this?

@typesafeai Opus 5.5 and my custom skill! I will release that soon as well

@typesafeai the 40% only holds while the agent keeps calling jg. skill instructions are the kind of context that compacts away first, and once they are gone runs quietly revert to raw greps. track adoption over long sessions, and re-inject the skill at compaction.

@typesafeai Ha, made one too -

@typesafeai This video is impressive wtf

@typesafeai How does this compare to FFF-grep?

@typesafeai I haven't benchmarked against the 2 but if there's enough interest I will (and against ripgrep as well)

@typesafeai Would be nice to see! Currently i’m using FFF almost everywhere’s supported

@typesafeai the skill so it actually uses the cheap tool. needed that

@typesafeai ngl the video looks great. I might give it a try at some point

@typesafeai sick!

@typesafeai just tried it, sadly not useful at all, super slow on repos, thx anyway!

@typesafeai How is this better than semble?

@typesafeai This is wild!

@typesafeai can u compare it against fff ?

@typesafeai @theo what about this?

A/B'd jg against ripgrep on a codebase we maintain. The right file often ranked first, but we still got a 12-51 file flood in seconds vs tens of ms for rg. NL ranking wins as the seed when you don't know the symbol, then you walk defs/refs/deps. Where does jg win beyond cold-repo exploration?

@typesafeai Do you think is good for Hermes?

@typesafeai cutting 40% token cost on swe-bench context collection is actually massive

@typesafeai This is definitely one of the smartest use cases of Jev I've seen

@typesafeai Worth a try, context is still a hard question right now

@typesafeai context collection is where the real agent cost hides — everyone optimizes the model call, nobody optimizes what feeds it. does the 40% hold pass@1 though, or did swe-bench resolution rate move with it?

@typesafeai what in tarnation is this motion design

@typesafeai Nice, a research CLI that cuts agent cost by 40% is super handy. The built-in skill for jg is a smart touch too. Congrats on the SWE-bench numbers!