Model router for coding agents

Every prompt, routed to the right model.

One endpoint and one model name, jev-auto, for Claude Code, Codex and Cursor. Renames go to a fast model; race conditions go to a frontier one.

$npm install -g jevrouter

$jev serve && jev setup

Step-by-step setup →Deploy free on Cloudflare

0 dependencies~30 ms startup9 providersMIT licence

POST /v1/messages

Prompt

Task: hard_debugging
Tier: frontier
Model: anthropic/claude-opus-5-5

x-jev-route: anthropic/claude-opus-5-5; task=hard_debugging; tier=3

Runs natively on Anthropic's own API, so the stream passes straight through.

How a request moves

Your agent talks to JEV. JEV picks the model.

JEV routes with a few rules that run in microseconds, with no extra LLM call. When the model speaks the client's protocol, bytes stream through untouched. When it doesn't, JEV translates both ways, including tool calls.

Claude CodeAnthropic Messages

CodexOpenAI Responses

CursorOpenAI Chat

jev-auto

ConstraintsKey present, context fits, supports tools and vision
ClassifyRename, debug, refactor, review… sets the tier
RankClosest tier, native protocol first, then price
StickTool loops stay on one model, which keeps caches warm
Recover429 or 5xx falls back; repeated tool errors move up a tier

Anthropic

OpenAI

Gemini

DeepSeek

Groq

Cerebras

OpenRouter

Mistral

Ollamalocal models, no data leaves your machine

Setup

Two commands, then use your tools as usual.

jev setup finds your clients, backs up their config and points them at JEV. jev uninstall-config restores every file exactly.

Install and add a key
Any provider works. Gemini and Groq have free tiers.
Start JEV
jev serve listens on 127.0.0.1:4000.
Run jev setup
This writes six env entries into ~/.claude/settings.json. Then run claude as normal.

~/.claude/settings.json

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://127.0.0.1:4000",
    "ANTHROPIC_AUTH_TOKEN": "<your JEV key>",
    "ANTHROPIC_MODEL": "jev-auto",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "jev-auto",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "jev-smart",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "jev-fast"
  }
}

Start JEV
jev serve
Run jev setup
This adds a marked provider block that uses the Responses wire format. An existing model line is commented out, not deleted.
Run codex
To try one model directly, run codex -m jev-smart.

~/.codex/config.toml

model = "jev-auto"
model_provider = "jev"

[model_providers.jev]
name = "JEV"
base_url = "http://127.0.0.1:4000/v1"
wire_api = "responses"

Deploy JEV to a public URL
Cursor sends requests from its own servers, so it can't reach localhost. Deploy free on Cloudflare.
Paste three values
Go to Settings → Models → API Keys. jev setup copies them to your clipboard.
Choose jev-auto
Pick it in the Chat or Agent model menu.

Cursor → Settings → Models

OpenAI API Key             <your JEV key>
Override OpenAI Base URL   https://jev.<you>.workers.dev/v1
Custom model               jev-auto

Not routed: Cursor Tab and other Cursor-owned features always use Cursor's own models. JEV handles Chat and Agent requests.

Use the JEV URL as the base URL
The three endpoints are /v1/messages, /v1/responses and /v1/chat/completions.
Use a JEV model name
jev-auto, or any exact model such as openai/gpt-5.
Read the route
Every response carries an x-jev-route header, and jev why explains the decision.

curl

curl http://127.0.0.1:4000/v1/chat/completions \
  -H "Authorization: Bearer $JEV_API_KEY" \
  -H "content-type: application/json" \
  -d '{"model":"jev-auto","messages":[{"role":"user","content":"Explain this regex"}]}'

Model names

Five names that work in every client.

Set it once. Change behaviour by changing the name, not the provider.

Name	What it does	Use it for
`jev-auto`	Classifies each prompt and picks the tier	Everything (the default)
`jev-fast`	Fast tier only	Titles, small edits, quick answers
`jev-smart`	Frontier tier only	Hard bugs, architecture, security work
`jev-cheap`	Lowest-priced available model	Bulk or background jobs
`jev-free`	Free-tier and local providers only	Zero spend; with Ollama alone, fully offline

Compatibility

What works today.

Feature	Claude Code	Codex	Cursor
Routing through JEV	Yes	Yes	Chat and Agent
Automatic setup	Yes	Yes	Paste 3 values
Streaming and tool calls	Yes	Yes	Yes
Fallback and escalation	Yes	Yes	Yes
Works with local `jev serve`	Yes	Yes	Needs a public URL
Tab completion	n/a	n/a	Handled by Cursor

MCP toolkit

Turn an API into tools your agent can use.

Start from a template or an OpenAPI spec. Generated servers are one file with no dependencies, and they run free on Cloudflare Workers.

jev mcp import openapi.json
Each operation becomes a tool. DELETE, admin and internal endpoints are skipped unless you pass --all.
jev mcp test .
Runs the handshake, checks schemas and calls every read-only tool, all without starting a server.
jev mcp lint .
Scores names and descriptions, marks tools read-only, write or destructive, and blocks committed secrets.
jev mcp deploy
Gives you a free public URL on workers.dev.
jev mcp install pets <url>
Adds the server to every client it finds and prints a one-click Cursor link.
Claude CodeCodexCursor

jev mcp import petstore.jsonreal output

$ jev mcp import petstore.json

  4 endpoints → 2 tools
  [x] get_pet_by_id         GET /pet/{petId}
  [x] find_pets_by_status   GET /pet/findByStatus
  [ ] DELETE /pet/{petId}   skipped: destructive
  [ ] POST /admin/flush     skipped: internal

$ jev mcp test pet-store-mcp

  ✓ initialize → pet-store-mcp 0.1.0
  ✓ tools/list → 2 tools
  ✓ get_pet_by_id: schema valid
  ✓ get_pet_by_id({"petId":1}) → 1423 ms
  ⚠ find_pets_by_status({}) → 400 missing `status`

  8 passed, 1 warnings, 0 failed

$ jev mcp lint pet-store-mcp

  MCP tool quality: 82/100
  ⚠ get_pet_by_id   READ_ONLY   parameters without description: petId

Free cloud

Run it in the cloud for $0.

The same code runs on Cloudflare Workers' free plan, which allows 100,000 requests a day with no cold starts. JEV refuses every request until you set a key, so it never runs as an open proxy.

Deploy

npx wrangler login
jev deploy                                   # prints your JEV key
npx wrangler secret put GEMINI_API_KEY       # once per provider key
jev setup --url https://jev.<you>.workers.dev --key <your JEV key>

FreeHosting

Cloudflare Workers is the main target. Deno Deploy and Vercel Edge run the same file.

Free tiersModels

Gemini, Groq, Cerebras and OpenRouter :free models. Use jev-free to stay on them.

Always freeA VM for local models

Oracle Cloud's ARM VM can run small Ollama models on its CPU.

No good optionGPUs

There's no reliable free GPU for serving models. Colab and Kaggle forbid it, so use the free APIs instead.

Stop choosing models by hand.

Install JEV, run jev setup, and keep working.

Get started Read the README