Give your AI agents autonomy.
Not unlimited access.

Jevrail evaluates every command before it runs, blocking dangerous actions and asking when things get risky.

go install github.com/SamanPandey-in/jevrail/cmd/jevrail@latest
Full installation guideShare on X
3 verdicts: allow, ask, deny.
DECISION
0 commands run past a deny.
SAFETY
1 policy layer for every agent.
CONTROL
∞ ways to keep shipping safely.
AUTONOMY

// PLATFORM

A safety layer for
agentic software.

Jevrail sits between an agent's decision and the shell. Every command is scored, explained, and resolved as allow, ask, or deny.

  ┌───────┐
  │ ◉ ◎ ○ │
  │ ◎ ◉ ◎ │
  │ ○ ◎ ◉ │
  └───────┘

Pre-execution analysis

Inspect commands before execution and surface the risk factors that shaped each verdict.

  ┌─┐   ┌─┐
  │A├──►│B│
  └─┘   └┬┘
        ┌▼┐
        │C│
        └─┘

Hard stops for harm

Stop destructive operations before they reach the shell, even when an agent is moving fast.

   ╔═══╗
   ║ ◈ ║
  ┌╨───╨┐
  │░░░░░│
  └─────┘

Context-aware decisions

Evaluate intent, environment, and command shape instead of relying on brittle string matching.

  │  ▄ 
  │▁ ▄ █ ▄
  │█ █ █ █
  └────────

Explainable verdicts

Get a human-readable reason for every allow, ask, or deny decision in the execution path.

    .--.
   /    \
  | (  ) |
   \    /
    '--'

Degraded-mode safe

When policy or context is unavailable, fail closed and keep the agent from taking an unsafe shortcut.

  GET /api
  ────────►
  ◄────────
  { data }

Works with your agent

Drop into your existing terminal workflow with a single Go install and a small policy surface.

// COMMANDS

A small interface for safer execution.

Install Jevrail once, then use focused commands to connect hooks, inspect decisions, and keep a clear audit trail.

Jevrail command reference
CommandPurpose
jevrail hook <claude|codex|opencode>Hook target: reads agent JSON on stdin, writes a decision (exit 2 = hard block).
jevrail install / uninstallIdempotent hook setup for Claude Code or opencode, with automatic config backups.
jevrail explain "<cmd>"Shows every probability, context, and verdict for a command without running it.
jevrail log [-n N]Prints the last N audit entries from the local jsonl audit log.
jevrail eval <corpus.jsonl>Benchmarks recall, false-ask rate, tier0 baseline, latency, and flip rate.
jevrail exec -- <cmd>Guarded execution shim for hookless agents: prompts on ask, blocks on deny.
jevrail doctorChecks config, API key, reachability, hook wiring, model pin, and timeout.
jevrail configure [--key KEY]Stores the Jev API key once; hook, explain, eval, and exec all reuse it.