jev-model-router 0.1.0

Quick start · How it works · User guide · Architecture · Contributing

JEV Router picks the Claude model for each new turn of your Claude Code session. Renaming a variable doesn't need Opus; debugging a race condition does. JEV Router asks TypeSafe's Jev — a fast decision model — which of your models fits the turn, then runs it there. Claude Code keeps working exactly as usual: same UI, tools, permissions, sessions and streaming.

pip install jev-model-router
jev-claude          # that's it — Claude Code, with per-turn routing

✨ Features

• Per-turn routing — every new message is judged on its own, so a session can move from Haiku to Opus and back as the work changes.
• Native Claude Code — a local proxy in front of Claude Code's API; nothing to configure in Claude Code, no fork, no plugin. /resume, permissions, tools and streaming are untouched.
• You stay in control — pick a model in /model and routing steps aside; pick JEV Router again to resume. Say "use opus" in a prompt to force it for one turn.
• Tool loops stay on one model — Jev is asked once per turn; every file read, edit and command in that turn reuses the chosen model.
• Fails open — no key, a Jev timeout or an error never blocks you; Claude Code carries on with a safe model within ~3 seconds.
• Explainable — a status-line readout of every decision, jev-explain for the full reasoning, and a one-line log per turn.
• Private by design — your key and prompts are never logged; local files are owner-only.
• Zero dependencies — pure Python standard library.

🎬 Demo

_{Left: "JEV Router" in Claude Code's own /model picker. Right: a rename routed to Haiku — the status line shows the model Jev picked.}

🚀 Quick start

Requirements: Python 3.11+, Claude Code installed and logged in (a claude.ai subscription or an API key — no extra Anthropic key needed), and a TypeSafe API key.

1. Install

pip install jev-model-router

_{Latest development version: pip install git+https://github.com/himanshu231204/jev_model_routers}

2. Add your TypeSafe key — TYPESAFE_API_KEY, the name TypeSafe's docs and SDK use:

echo "TYPESAFE_API_KEY=your_key" > ~/.jev-router.env        # or: export TYPESAFE_API_KEY=...

_{Windows PowerShell: Set-Content "$HOME.jev-router.env" "TYPESAFE_API_KEY=your_key"}

3. Run

jev-claude                          # interactive, routing each new turn
jev-claude -p "fix the failing test"   # every Claude Code argument is passed through
jev-claude --resume                 # sessions work as usual

The session starts on JEV Router. Watch the status line — e.g. claude-haiku-4-5-20251001 (p=0.97) — to see which model each turn got. The full guide, with troubleshooting, is in docs/quickstart.md.

🧠 How it works

flowchart LR
    A["Claude Code"] -->|"model = jev-router"| B["JEV Router<br/>(local proxy)"]
    B -->|"new turn: ask Jev"| J["Jev<br/>(TypeSafe System One)"]
    J -->|"best model + confidence"| B
    B -->|"model = claude-haiku / sonnet / opus"| C["Anthropic API"]
    C -->|"streamed back, unchanged"| B --> A

1. jev-claude starts Claude Code with a local proxy as its API endpoint and a JEV Router entry in /model.
2. When you send a message, the proxy asks Jev which of your account's models — newest Haiku, Sonnet or Opus — can handle it, with a confidence score.
3. A small local policy makes the final call: explicit choices win, low confidence never downgrades, and a long conversation isn't downgraded (switching would re-read it all).
4. Only the request's model is rewritten; Claude's response streams straight back.
5. Tool calls, retries and Claude Code's background requests in the same turn reuse that model.

Jev doesn't write code — Claude does. Jev answers one quick structured question per turn, the kind of "route this to one of a fixed set of destinations, and know how confident you are" decision it is built for. Design details: ARCHITECTURE.md.

⚙️ Configuration

Everything works with just the key. Optional environment variables:

Routing thresholds (confidence floor, cache protection, timeouts) live in one file: src/jev_router_live/config.py.

🧰 Commands

❓ FAQ

📍 Project status

Beta. Verified end to end against Claude Code 2.1.282 (interactive and -p): the picker entry, per-turn routing, tool-loop pinning, manual override, fail-open and streaming, with 94 automated tests on Python 3.11 and 3.12. OpenAI Codex support (jev-codex) is experimental. Known limitations are listed in ARCHITECTURE.md.

🤝 Contributing

Contributions are welcome — see CONTRIBUTING.md for setup and conventions and AGENTS.md for the invariants a change must keep. Please report security issues privately as described in SECURITY.md.

git clone https://github.com/himanshu231204/jev_model_routers.git && cd jev_model_routers
pip install -e ".[test]"
python -m pytest -q

📄 License

MIT © 2026 Himanshu Kumar

_{JEV Router is an independent open-source project and is not affiliated with Anthropic or TypeSafe. Claude and Claude Code are products of Anthropic; Jev is a model by TypeSafe.}