User repository on GitHub
Jev Router for Cursor - TypeSafe Jev picks COST/BALANCED/INTELLIGENCE, Cursor executes
- Forks
- 1
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Jev-chooses-a-LLM
Cursor router that has Jev pick a cost, balanced or intelligence tier per task, then enforces deterministic policy on top: security-sensitive work escalates and financial or deploy actions need approval.
- Stars
- 4
- Language
- TypeScript
- License
- MIT
- Published by
- Community
- Author
- Bodila51
- Last verified
- Oct 2026
About this project
Cursor executes. LLMs write and build. Jev decides how much intelligence the task deserves. Honest promise: pick the cheapest available route likely to complete successfully under policy — not a mathematically optimal model. TypeSafe Jev is a decision layer only (Choice / Score / Noul). It does not browse, write prose, edit files, or operate accounts. Env: TYPESAFEAPIKEY. In the Cursor Agents window for this project, paste:
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- TypeScript
Highlights
- Chat mode is soft: the parent calls jev_route then delegates to jev-cost, jev-balanced or jev-intelligence subagents.
- SDK mode is hard: route, validate, discover models, then Agent.create and send with the resolved catalog model, with dry-run.
- Confidence below 0.70 routes to BALANCED, security-sensitive tasks get at least INTELLIGENCE, financial and deploy actions need approval.
- COST failure escalates to BALANCED then INTELLIGENCE then stops, capped at two escalations and no same-tier retry.
Watch out
The README is blunt about the target: the cheapest available route likely to complete under policy, not a mathematically optimal model. The honest limit is that chat mode cannot force the parent model ID, so Jev picks the tier and the parent delegates; only the SDK runner with --run performs an actual model switch. Live SDK runs need CURSOR_API_KEY and Node 22.13 or newer for @cursor/sdk, and Cursor Router's auto-smart mode is account-dependent, so the runner falls back to catalog model IDs and never invents them. Approval gates cover send_email, delete_data, publish, production_deploy and financial_transaction, and stop SDK runs unless you pass --approve. If Jev is unavailable the router selects BALANCED and sets degradedRouting, and with no TYPESAFE_API_KEY the dry run and demo use a deterministic mock. Cost is read from run.usage or agent.getUsage and labeled unavailable when the plan does not expose it. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
Only in SDK mode. The MCP chat path decides a tier and then delegates to a subagent for that tier, while the parent model keeps running, which the README calls out as the honest limit. Real switching happens through the SDK runner, where the resolved model feeds Agent.create and send, and that needs CURSOR_API_KEY plus Node 22.13 or newer.
Not through the SDK path without you saying so. Approval gates for send_email, delete_data, publish, production_deploy and financial_transaction stop a run unless you pass --approve, and there is a kill switch: enabled false in config or /jev-bypass for a single turn. The optional Gmail and X flows are documented as patterns only, with no bundled OAuth apps.
Live routing falls back to BALANCED and flags degradedRouting so the run is not silently pretending the decision came from the model. With no TYPESAFE_API_KEY the dry run and demo use a deterministic mock, which is fine for wiring up config but tells you nothing about real routing quality. Log the routing trace while you evaluate the policy.
None in the repository. The test suite runs against a mock TypeSafe client and eleven product scenarios with no real keys, which tests that the policy branches fire, not that routing improves outcomes. No benchmark compares cost, latency or success rate across tiers, and cost is only reported when the plan exposes billed cents. Judge the policy rules yourself and set maximumTaskCostUsd accordingly.
Clone the repo, copy .env.example to .env and set TYPESAFE_API_KEY (plus CURSOR_API_KEY for live SDK runs), then npm install, npm test and npm run demo. Chat mode needs Node 20 or newer and the plugin symlinked into ~/.cursor/plugins/local with MCP jev-router enabled; live SDK runs need Node 22.13 or newer for @cursor/sdk.
Confidence below 0.70 selects BALANCED; security-sensitive work gets at least INTELLIGENCE; financial or production-deploy actions get INTELLIGENCE plus a human approval flag; send or delete email and publish also need approval; read-only research is fine; a COST failure escalates to BALANCED then INTELLIGENCE then stops, capped at two escalations with no same-tier retry.
config.example.json holds the default mode (balanced), the 0.7 confidence threshold, maximum escalations (2), a maximum task cost of $1.00, routing-trace and prompt-logging toggles, and the requireApprovalFor list covering send_email, delete_data, publish, production_deploy and financial_transaction. A kill switch disables it for a turn or entirely.
More Jev resources like this
20GitHub stars
TypeScript SDK and CLI coding agent where Jev only picks options while your code runs effects, building programs as syntax trees.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
rhighs/jev-codeTypeScript
rhighs/jev-codeTypeScript13GitHub stars
Pi extension that classifies each task with Jev over OpenRouter and routes it to a Pareto-optimal model balancing measured quality, cost, and latency under a role policy.
Jev Agents & MCP#community#jevbest#catalog-import
12GitHub stars
Local OpenAI-compatible proxy that skips repeat model calls on same-intent paraphrases and retried agent turns, judged by Jev or a local System One model.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
kushals256/jevcacheTypeScript
kushals256/jevcacheTypeScript10GitHub stars
Node router that sends each query to the cheapest capable LLM tier by asking Jev to classify demand against your tier descriptions, with no LLM call spent on routing.
Jev Agents & MCP#community#jevbest#catalog-import
Community
rajdhakad9826/jev-routerTypeScript
rajdhakad9826/jev-routerTypeScript7GitHub stars
Claude Code plugin that scores pending tool calls and finished answers with Jev noul probabilities, denying the call with the correction attached when a rule threshold is met.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
muratcakmak/jev-guardTypeScript
muratcakmak/jev-guardTypeScript6GitHub stars
Terminal coding agent on Pi where Jev shortlists skills, ranks retrieved excerpts, and triages tool failures while the coding model handles reasoning and edits.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
tonyzdev/pijevTypeScript
tonyzdev/pijevTypeScriptFrom the Jev community
Posts from builders shipping with Jev right now.