Quick start

uv tool install jevpipe                             # the CLI
jevpipe auth set-key                                # your API key, kept in the system keychain
npx skills add fabianboth/jevpipe --skill jevpipe   # teaches your agent when and how to use it

You need an OpenRouter API key, or a TypeSafe API key after jevpipe config set provider typesafe. Where there is no keychain, as in containers, CI or headless Linux, set OPENROUTER_API_KEY or TYPESAFE_API_KEY instead; it also takes precedence. No uv yet? Install it or use pipx.

pipx install jevpipe                                # instead of uv
gh skill install fabianboth/jevpipe jevpipe         # the skill, with the GitHub CLI

Example

Which of the 19 commits in ripgrep 15.1 are new features? Ask each commit message:

# in a clone of https://github.com/BurntSushi/ripgrep
git log --format=%s 15.0.0..15.1.0 | jevpipe filter "Is this a new feature?"

# the commit messages that match the question
ignore/types: add `ssa` type
printer: add Cursor hyperlink alias

# the summary, on standard error
jevpipe: 19 records, 2 kept, 0 skipped, 0 failed, 5.7k tokens, $0.000225, 1.1s

# in a clone of https://github.com/BurntSushi/ripgrep
git log --format=%s 15.0.0..15.1.0 | jevpipe map -q '{
  "kind": {
    "type": "choice",
    "instructions": "What kind of change is this, for the release notes?",
    "criteria": {
      "feature": "a new capability for users",
      "fix": "a bug fix users would notice",
      "docs": "documentation only",
      "internal": "refactoring, tests, CI, dependencies or release chores"
    }
  }
}'

{
  "record": "printer: add Cursor hyperlink alias",
  "answers": {
    "kind": {
      "type": "choice",
      "choice": "feature",
      "probabilities": {"docs": 0.05, "feature": 0.87, "fix": 0.03, "internal": 0.05},
      "confidence": 0.82
    }
  }
}

Use cases

Each item is judged on its own, so ask what the item itself can answer:

Measured

Classification. Jev against general LLMs on intent routing and prompt-injection detection, from an independent study:

Code search. jevpipe in an agent's hands: 470 searches from the extended CodeSearchNet Challenge in five languages, each over every function of its language.

Read the full benchmark: every language, speed and cost, and the searches where jevpipe loses.

Commands

Both read their input like grep: lines through a pipe or from the files you name. With --read-files, each line is a path, and the file it names is judged instead:

git log --format=%s | jevpipe filter "Is this a new feature?"                       # each commit message
jevpipe filter "Is this an error worth a closer look?" app.log                      # each line of app.log
git ls-files | jevpipe filter "Does this file retry failed requests?" --read-files  # each file

jevpipe <command> --help has the rest.

Cost and speed

Every record is one request billed to your OpenRouter or TypeSafe credit, and the cost follows the size of each record: about a cent per 1,000 commit messages, 4 to 18 cents per 1,000 source files. Asking several questions at once barely changes it. Up to 100 records run at the same time, so hundreds take seconds.

--max-cost 0.50 stops sending requests once the run has spent $0.50; it then exits with status 3 and names the line to resume from. On TypeSafe, which reports tokens rather than dollars, use --max-tokens 5M instead. To cap every run by default, run jevpipe config set max-cost 0.50 once; jevpipe config --help lists the other defaults you can set.

License

Apache-2.0. jevpipe is an independent project, not affiliated with or endorsed by TypeSafe AI; the answers come from Jev, TypeSafe AI's calibrated decision model, through OpenRouter or TypeSafe's own API. Building from source needs Rust (rustup picks the pinned toolchain) and PowerShell 7 for ./check.ps1.