Jev directory
CommunityJev Tools & CLIs
Jev Compiler (Jevc)
A CLI that compiles natural-language agent rules into Jev programs: a few narrow typed questions answered by TypeSafe's System One model plus a reducer that computes the verdict in ordinary code, so a markdown rule becomes an invariant you can read, diff and unit-test.
- Reading time
- 2 min
- Published by
- Community
- Added
- Nov 2026
Highlights
- Compiles natural-language rules into typed Jev questions plus a reducer written in ordinary code.
- jevc scan reads CLAUDE.md, AGENTS.md, GEMINI.md, .cursorrules, copilot-instructions and .claude skill and agent files.
- Ships 58 recorded fixtures; jevc check replays them offline with no API key and no network.
- jevc compile --lift prints a lowering request for an agent you already have open, with no second API key or bill.
- Targets the non-generative jev-1.13.0 model, which returns probabilities and never text; Node >= 22 runs the test suite.
Questions about this entry
Short answers drawn from the linked project or page.
Install globally with npm install -g jev-compiler; the package is jev-compiler and the command it installs is jevc. From source, clone the repository, run npm install and npm run build, then use npx jevc from the checkout.
No, not for the offline surfaces. The 58 recorded fixtures are replayed by jevc check with no API key and no network, and the examples run the same way. Only jevc check --live reaches the model.
It finds the files agent harnesses actually read: CLAUDE.md, AGENTS.md, GEMINI.md, .cursorrules, .github/copilot-instructions.md, .claude/skills/** and .claude/agents/**.
Decisions such as is this command dangerous or which team owns this, which compile into typed questions. Generation never compiles because Jev emits no text, and procedure such as read the file before editing belongs in code as control flow. A compilation that produces no decisions is a valid answer.
TypeSafe's System One model, jev-1.13.0 in the published fixtures. It is non-generative: it returns probabilities and never text, so the verdict is computed by the reducer rather than written by a model.
Every compilation emits all three parts: the decision set, a residual prompt for whatever genuinely still needs a generative model, and the wiring that evaluates the decisions and calls the LLM only for the residual.
It is a text heuristic and the project says so itself. Because of that, the list of rules it flags is worth reading rather than trusting the counts, and not every rule it reports as decidable is worth compiling.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Vercel's fx safety reviewer, measured at 18× faster
Reddit: Early experiments wiring Jev into Pi as a safety layer and model router.
YouTube: Ray Amjad puts Jev inside an agentic coding loop and looks at what it costs to run.
YouTube: RoboNuggets puts Jev to work inside Claude Code.
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.