User repository on GitHub
Route each Codex turn to the cheapest model that can handle it, judged by TypeSafe Jev.
- Forks
- 2
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Codex Sift
A model router for the Codex CLI: on every user turn Jev judges how hard the task is and a YAML policy picks one of three lanes — flash, craft or forge — so premium quota goes to the prompts that need it.
- Stars
- 8
- Language
- TypeScript
- License
- MIT
- Published by
- Community
- Author
- romanmeclazcke
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Languages
- TypeScript
- Platforms
- CLI
- Capabilities
- Open sourceRequires API key
Highlights
- Lanes map to models: flash gpt-5.6-luna, craft gpt-5.6-terra, forge gpt-5.6-sol, with the first matching rule in policy.yaml winning.
- Jev answers atomic signals such as task_kind, difficulty, needs_planning and high_stakes; the policy never asks it which model to use.
- E2E report: 12 cases routed live at 418 ms average, 11 of 12 matching policy expectations, 53.7% estimated quota saved against always-forge.
- A loopback Responses proxy intercepts POST /responses through a custom model_provider and forwards your existing login upstream.
- codex-sift explain and the $sift-explain skill print the last decision from ~/.codex-sift/last.json after any routed turn.
Watch out
MIT, and Codex-only: it runs a loopback Responses proxy and is not an MCP server, so Codex starts through its own command and another model pinned in /model pauses routing. Live routing needs a TYPESAFE_API_KEY; without one, or if Jev times out, it fail-opens to the craft lane and marks the decision degraded.
Questions about this entry
Short answers drawn from the linked project or page.
Coding sessions mix tiny asks with genuinely hard work; running everything on the strongest model spends premium quota on prompts like "run the tests". Sift decides a lane per user turn while you keep the native Codex TUI.
Codex has no hook that can change the model before a turn is billed, so Sift starts a loopback Responses proxy, launches Codex with process-local -c overrides, and forwards your existing ChatGPT or API-key login upstream.
The latest E2E report covers 12 representative cases, all routed with live Jev, none degraded, 11 of 12 matching policy expectations, average Jev latency of 418 ms, and estimated quota saved of 53.7% against an always-forge baseline.
It fail-opens: the turn runs on the on_jev_down lane, craft by default, and the decision is marked degraded so the savings report is not inflated by failed calls.
Tool-loop continuations keep the chosen model pinned, codex-auto-review and compact or review prompts stay native, housekeeping that is not a /responses call is untouched, and anything launched with -m or --model bypasses routing; --sift-off bypasses the proxy for one invocation.
npm install, npm run build, npm link, then codex-sift setup lists your Codex models as numbers and you pick one for flash, craft and forge; it can save TYPESAFE_API_KEY to ~/.codex-sift/env at mode 600, and doctor checks the binary, auth mode and key.
No. Codex itself sends the Authorization header and, on ChatGPT login, chatgpt-account-id; the proxy forwards them to chatgpt.com/backend-api/codex or api.openai.com, and the README warns a ChatGPT token is not valid on the platform API.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Facebook: Intake Router keeps diagnosis out of the answer space at the front desk.
FacebookClinicnotes
YouTube: Jev explained for routing and judging, with a comparison against LLMs.
Hacker News: Vercel on using Jev for defined decisions and GPT-6 Astra when the job also has to write or use tools.
flashbrew4 pointsTypeSafe AI Jev vs. GPT-6 AstraRead the thread on Hacker NewsHacker News: ASTGL replay chose local Laya routing over hosted Jev on one Mac Studio setup.
Jmeg8r15 pointsDoes an open-weight decision model beat a hosted one? Jev vs. LayaRead the thread on Hacker NewsMore Jev resources like this
32.6kGitHub stars
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-templatesPython7GitHub stars
Launchers for Claude Code, Codex, Grok CLI, and opencode that ask Jev which model-and-effort tier each human turn needs, rewriting only the model field while the real CLI runs unchanged.
Jev Agents & MCP#community#javascript#routing
Community
flaviusapop/jev-routerJavaScript
flaviusapop/jev-routerJavaScript35.4kGitHub stars
Stencil Labs' coding agent embeds TypeSafe Jev for typed checks, routing, and workflow decisions, with measured Jev tokenizer probes in pi-natives.
Jev Agents & MCP#community#typescript#coding-agent
Community
can1357/oh-my-piTypeScript
can1357/oh-my-piTypeScript59GitHub stars
A free Claude Code skill that puts Jev in front of the expensive model: thirteen commands sort, check, score, gate and verify with Jev, so Claude reads only the unsure slice and writes the output. Needs python3 and an OpenRouter key.
Jev Agents & MCP#community#python#claude-code
Community
charlesdove977/claude-x-jevPython
charlesdove977/claude-x-jevPython45GitHub stars
Five typed tools — classify, check, score, rank, and ask — plus a skill, wired into Claude Code, Codex, Pi, and OpenCode in one setup command.
Jev Agents & MCP#community#typescript#mcp
Community
FrancoisChastel/jev-codeTypeScript
FrancoisChastel/jev-codeTypeScript8GitHub stars
A 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-pluginTypeScriptFrom the Jev community
Posts from builders shipping with Jev right now.