Jev directory
CommunityJev Agents & MCP
@riposta/pi-jev
A typed classification layer for the Pi coding agent: five independently switchable Jev modules — router, gate, shield, prune, and watchdog — that ship in shadow mode, one request per hook, with measured fixture numbers.
- Reading time
- 3 min
- Published by
- Community
- Author
- riposta
- Last verified
- Oct 2026
- Languages
- TypeScript
- Platforms
- CLI
- Capabilities
- Open sourceRequires API key
Highlights
- Five switchable modules — router, gate, shield, prune, and watchdog — each making one request per hook.
- Measured fixtures: router tier accuracy 90-92%, gate 0/19 false negatives on dangerous commands, shield 93.3% injection detection.
- Requires Pi 0.85.0+ and Node 22.19+; install the scoped @riposta/pi-jev, not the unrelated unscoped pi-jev.
- Every module ships in shadow mode and changes nothing until promoted with /jev shadow or config.
- QUESTIONS_VERSION hashes questions.ts, so edits invalidate the memory cache, disk cache, and calibration generations.
Watch out
The README's numbers come from the project's own v0 fixtures, not a vendor benchmark, and router net cost, gate cache hit rates, and prune token savings are explicitly not yet measured. The extension sends metadata and sampled tool output to api.typesafe.ai, so read the privacy section before enabling it on sensitive repositories.
Questions about this entry
Short answers drawn from the linked project or page.
A typed classification layer for the Pi coding agent backed by TypeSafe's jev-latest model. It turns implicit agent decisions — model tier, command recoverability, log value, task completion — into explicit Jev questions with probabilities, confidence, and reviewable thresholds, across five modules.
Pi comes first: install it (Node >= 22.19) and set TYPESAFE_API_KEY or /login typesafe. Then run pi install npm:@riposta/pi-jev with Pi 0.85.0+ — use the scoped name, since the unscoped pi-jev belongs to an unrelated project — or pi install git:github.com/riposta/pi-jev.
From the repository's labelled fixtures against real api.typesafe.ai: router tier accuracy 90-92%, added latency p50 288-673 ms, gate false negatives 0/19, gate false positives 5.0%, shield injection detection 93.3%, and zero failures that block Pi. The README invites you to re-run tools/evaluate.ts.
Every module classifies and logs the counterfactual decision but changes nothing until promoted. Promote one for a session with /jev shadow gate off, or permanently by setting modules.<name>.shadow or enabled in ~/.pi/agent/jev.json or a trusted project's .pi/jev.json.
The key comes from TYPESAFE_API_KEY or Pi's /login typesafe flow, with the environment variable winning when present; check state with /jev auth. The extension sends metadata and sampled tool output to api.typesafe.ai, and there is no phone-home beyond those requests.
/jev for status and usage, /jev auth for key state, /jev explain for the last classification, /jev shadow to toggle a module, /jev off and /jev on as session kill switches, /jev stats for the decision mix, /jev trace for recent decisions, and /jev recommend for threshold advice.
Router net cost change (needs model pricing), gate disk-cache hit rate over real sessions (needs a warm log), and prune token saving. The README lists these as open items and notes the fixtures are a v0 seed.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Agent Handoff Gate: agents verify what they hand off
X: Jev as an LLM guardrail inside Agentgateway
X: Vercel's fx safety reviewer, measured at 18× faster
Reddit: Early experiments wiring Jev into Pi as a safety layer and model router.
More Jev resources like this
A Claude Code configuration CLI shipping Jev modules for model routing, skill suggestion, tool guardrails, auto-mode gates, and sandbox hooks.
Jev Agents & MCP#community#typescript#coding-agent
Community
davila7/claude-code-templatesPython
davila7/claude-code-templatesPythonA Pi extension that reviews shell commands before execution in three tiers — auto-permitted, auto-blocked, and model-reviewed — with Jev, OpenRouter, or a pi model returning ALLOW or BLOCK for risky commands.
Jev Agents & MCP#community#typescript#pi
Community
An MCP proxy and hooks adapter that screens an agent's tool calls before they run and the tool results before the agent reads them, with thresholds in a policy file.
Jev Agents & MCP#community#mcp#security
A guard for coding agents that puts Jev in front of every step: it judges tool calls for intent and risk, scans tool outputs for instructions aimed at AI agents, and checks replies, mapping probabilities to allow, ask or warn.
Jev Agents & MCP#community#catalog-import#logicrw
Community
harshwasan/pi-jev-sentinelTypeScript
harshwasan/pi-jev-sentinelTypeScriptA DeepSeek Harness plugin that exposes Jev's noul, choice and score primitives as agent tools, plus configurable guardrails, retries, response caching, a circuit breaker and Web UI result cards.
Jev Agents & MCP#community#typescript#deepseek
Community
jackie-cqz/dsh-jev-pluginTypeScript
jackie-cqz/dsh-jev-pluginTypeScriptA Claude Code plugin that hands Claude Code's small, frequent decisions to Jev at six points: four event hooks and two routing decisions, each a roughly one-second call that reduces permission prompts, warns when fetched content addresses the agent, re-checks done, and routes effort and model.
Jev Agents & MCP#community#github#claude-code
Community
yinjs/claude-jevTypeScript
yinjs/claude-jevTypeScriptFrom 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.