claude-router

A lightweight local proxy that sits between Claude Code and LM Studio, enabling fully local, zero-cloud agentic coding with automatic model routing, enabling usage of a supporting vision capable model.

• Text/code requests → qwen/qwen3-coder-next (primary coder)
• Image/vision requests → mistralai/devstral-small-2-2512 (vision + coding)

Once images are analyzed and cached, all subsequent turns automatically route back to the coder model — with image descriptions injected as text so the coder retains full visual context without ever receiving raw image data.

NOTE LIMITATIONS BELOW

IMPORTANT: Image description cache is in-memory — lost when the router restarts! Start a new Claude Code session after restarting the router! Or tell it to recheck the images as they have "changed".

---

How It Works

Claude Code
    │
    ▼
claude-router :3000
    ├── last user message has uncached images? ──► Devstral (vision)
    │                                                   │
    │                                         caches description
    │                                                   │
    └── text / cached images ◄──────────────────────────┘
              │
              ▼ (images replaced with <<IMAGE DESCRIPTION: ...>>)
         Qwen3 Coder Next
              │
              ▼
         LM Studio :1234
              │
              ▼
         your machine
         zero cloud

Message Transforms

The router applies several transforms to work around prompt template limitations of local models:

Transform	Purpose
rewriteForVision	Extracts images from tool_result blocks, injects synthetic assistant/user turn
substituteImagesWithDescriptions	Replaces raw images in coder-bound requests with cached text descriptions
normalizeAssistantMessages	Merges interleaved text+tool_use blocks (required by Devstral's jinja template)
fixMessageAlternation	Merges consecutive same-role messages to enforce strict alternation
Strip thinking / betas	Removes Claude-specific params unsupported by local models

---

Requirements

• Node.js 18+
• LM Studio 0.4.1+ (for Anthropic-compatible /v1/messages endpoint)
• Claude Code installed globally
• Models loaded in LM Studio:
  • qwen/qwen3-coder-next
  • mistralai/devstral-small-2-2512

---

Installation

Comfy case, otherwise just run like described in Usage.2 "Start the router" (node ~/bin/claude-router.mjs)

# 1. Place the router in your personal bin
mkdir -p ~/bin
cp claude-router.mjs ~/bin/claude-router.mjs
chmod +x ~/bin/claude-router.mjs

# 2. Make sure ~/bin is on your PATH (add to ~/.zshrc if needed)
export PATH="$HOME/bin:$PATH"

---

Usage

1. Load both models in LM Studio

lms load qwen/qwen3-coder-next --gpu max
lms load mistralai/devstral-small-2-2512 --gpu max

# Verify both are running
lms ps

Or just do it in the GUI..

Tip: Devstral defaults to 4096 context. Bump it for longer sessions:

lms load mistralai/devstral-small-2-2512 --context-length 32768 --gpu max

2. Start the router

node ~/bin/claude-router.mjs

Expected output:

🔀 claude-router on :3000
   coder  → qwen/qwen3-coder-next
   vision → mistralai/devstral-small-2-2512

3. Run Claude Code via the router

CRUDE TO BE HONEST!

Create a launcher script ~/bin/claude-local:

#!/bin/bash
unset ANTHROPIC_API_KEY

ANTHROPIC_BASE_URL="http://127.0.0.1:3000" \
ANTHROPIC_AUTH_TOKEN="lmstudio" \
ANTHROPIC_API_KEY="" \
CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 \
  claude "$@"

chmod +x ~/bin/claude-local

Then use claude-local instead of claude for local sessions:

claude-local        # uses LM Studio via router
claude              # unchanged — still uses Anthropic servers

---

Configuration

Edit the constants at the top of claude-router.mjs:

const CODER_MODEL  = "qwen/qwen3-coder-next";           // model for code tasks
const VISION_MODEL = "mistralai/devstral-small-2-2512"; // model for image tasks
const PROXY_PORT   = 3000;                              // router port

LM Studio is assumed to be running on 127.0.0.1:1234. To change this, update the forward() function:

hostname: "127.0.0.1",
port: 1234,

---

Routing Logic

Condition	Model
Last user message contains uncached images	VISION_MODEL
Last user message contains only cached images	CODER_MODEL + descriptions injected
Last user message is text only	CODER_MODEL

Images are considered "cached" once Devstral has analyzed them and returned a text description. The cache persists for the lifetime of the router process (in-memory, per session).

---

Router Output

→ coder    [qwen/qwen3-coder-next]
→ coder    [qwen/qwen3-coder-next]
→ vision   [mistralai/devstral-small-2-2512]
  cached description for 3 image(s)
→ coder    [qwen/qwen3-coder-next]
→ coder    [qwen/qwen3-coder-next]

---

Tested Environment

Component	Version
LM Studio	0.4.10
Coder model	qwen/qwen3-coder-next (44.86 GB, 131072 ctx)
Vision model	mistralai/devstral-small-2-2512 (14.12 GB)
Node.js	25.9.0

---

Limitations

• Image description cache is in-memory — lost when the router restarts. Start a new Claude Code session after restarting the router.
• Streaming responses from Devstral may not be parseable as JSON for caching. If no description is cached, images fall back to <<IMAGE: binary image file>> in coder history.
• Not a production proxy — no TLS, no auth, no rate limiting. Intended for local single-user use only.
• Model name must match exactly what lms ps reports. A mismatch silently falls back to whatever LM Studio has loaded as default.

---

License

MIT