Open source · MIT · Built in Zig

A typed decision layer
for Codex.

jevx pairs Codex’s generative coding with Jev’s bounded choice, noul, and score judgments—then keeps final control in deterministic policy you can inspect.

Unofficial community project. Not endorsed by OpenAI, TypeSafe, or OpenRouter.

01

Typed routing

Structured semantic judgments instead of free-form permission guesses.

02

Deterministic gates

Thresholds and immutable guards remain ordinary, testable Zig code.

03

Encrypted evidence

Redacted, chained audit records make terminal work easier to verify.

One command. Guided setup.

The installer selects the correct release, verifies its checksum, keeps bundled notices intact, and launches jevx setup.

curl -fsSL https://raw.githubusercontent.com/SupratimSircar05/jev-zig-cli/v1.0.0/install.sh | sh
macOS Apple Silicon macOS Intel Linux arm64 Linux x86_64 Windows x86_64

The one-line installer supports macOS and Linux, including Linux under WSL. Native Windows users should download and verify the provenance-attested x86_64 archive directly. Platform binaries are currently OS-unsigned; verify GitHub provenance and SHA-256 before use.

See the decision boundary.

Connect to jevx web on your machine for live local decisions. Without it, the terminal uses a deterministic browser simulation that never calls Jev.

CHECKING LOCAL BRIDGE
jevx://decision-lab

The pairing token is an ephemeral local bridge secret, not a provider API key. It stays in this tab's memory and is cleared on exit. Live prompts go only to the loopback bridge on this device.

OUTPUT DETERMINISTIC BROWSER DEMO · NOT REAL JEV
Waiting for a decision…

Judgment is typed.
Authority stays in code.

Jev does not generate code here. It supplies bounded signals. Codex performs generative work only after jevx validates the response and applies its own policy.

CLOSED OUTPUTS

Small semantic surface

Only three validated decision types cross the Jev boundary. Choice lists and score ranges are bounded.

CODE OWNS CONTROL

Inspectable permission logic

Confidence, impact, and hazard thresholds are explicit. A model cannot waive immutable guards.

FAIL CLOSED

Safe degradation

If typed preflight or durable auditing fails, agent turns stop before Codex receives the prompt.

Useful autonomy,
with hard edges.

“Better” means more auditable and predictable—not smarter than the underlying model or a replacement for human judgment.

  • Always ask

    Pushes, releases, deployments, messages, purchases, account changes, destructive operations, privilege changes, credential access, and writes outside the workspace.

  • Always deny

    Credential exfiltration, sandbox bypass, policy tampering, and broad destructive commands.

  • Never hide

    Versioned JSONL events, stable exit classes, redacted encrypted audit records, and explicit degraded states.

  • Never loop

    One batched postflight verification and at most one repair turn for a mutating task.

Built for humans and scripts.

Stream in a terminal, compose through stdin, or consume stable JSONL.

jevx

Start the streaming interactive REPL.

jevx run

Run one Codex turn with the prompt on stdin.

jevx resume ID

Continue a bound Codex thread safely.

jevx decide

Submit one typed Jev request from stdin.

jevx web

Connect this site to the loopback-only typed decision bridge.

jevx doctor

Check dependencies, authentication, policy, and audit health.

jevx policy explain

Print effective thresholds and immutable guards.

jevx audit verify

Validate encrypted frames and chain integrity.

jevx --json

Emit versioned, machine-readable JSON Lines.

OPEN SOURCE · UNOFFICIAL · AUDITABLE

Put a decision boundary
around your coding agent.

Copied to clipboard