Dev tools built with Jev

Jev reads the prompt first and picks the skill, tool and parameters before the agent starts, halving its response time.

Jev looks at a PDF page by page and decides which pages need OCR; the rest is extracted locally. Cheaper and faster.

Jev chooses the join order; after some tuning, queries on the Join Order Benchmark ran 12% faster.

Jev picks the agent, model, computer and folder for each task: Fable with Claude Code for a big rewrite, a Mac for iOS changes.

No more nested if/then logic: Jev picks the next question, json-render draws it and xstate keeps the state.

As soon as you stop typing, Jev rates the prompt; if it's simple, the app offers fast mode.

A fork of nanocode with a searchable memory file, where Jev compaction drops the memories that no longer matter.

In the goose agent harness, Jev reads the prompt and selects the model right before each turn runs.

Same prompt and 15 test steps through Wikipedia: Jev beat DeepSeek with vision by 14.8x.

Jev scores every old tool output, keeps what's needed and parks the rest on disk: 30–55% less context for $0.0005 a pass.

json-render with Jev: interfaces assembled from your own components, actions and design system, rendered in milliseconds. Open source.

One call returns 14 typed checks on a diff — secrets, SQL injection, deleted tests, blast radius — and code turns them into block, review or merge.

Jev acts as the computer itself and runs very short programs, sometimes wrong against a reference emulator.

Jev reviews your row-level security policies and flags the risky ones.

Same test, same loop: Jev took 29 s and $0.003, Haiku 4.5 took 37 s and $0.07. Open source.

Prompt classification in ~140 ms decides which LLM gets each request: 20x cheaper and 6x faster than before.

A 27K-parameter model parses queries into filters in 0.25 ms in your browser, and asks Jev only about the words it's unsure of.

Give it your data and a target like 99% accuracy; it returns the exact threshold and how much still needs an LLM. Open source.

With ~90 skills installed and 5 in use, Jev picks the right skill for each request. Open source.

Test Choice, Score and Noul requests, mix them, and see whether your results improve. Seed examples included.

Traces show which tool ran, not whether there's progress. JevScope has Jev watch the agent and flag loops like editing the same file six times.

Filter vanilla Postgres by meaning with Jev. No extension, nothing to ask your DBA for.

Compaction for coding agents without a summarization prompt: Jev scores every tool call and drops what is no longer relevant.

WHERE jev(people, 'could work from home'): 129 rows judged in ~1 s for $0.0009, and 6 ms from cache the second time.