Skip to content

About

Connect Devin models (GLM-5.2, etc.) to Claude Code + ClaudeKit via 9router. No Anthropic API key needed.

Resources

Code of conduct

Contributing

Stars

4 stars

Watchers

0 watching

Forks

Repository files navigation

Devin → 9Router Bridge

Use Devin's GLM-5.2 (and other Cognition models) as a backend for Claude Code — no Anthropic API key required.

💡 Why GLM-5.2? Devin Pro (Windsurf) includes promo free credits for GLM-5.2 High, making it essentially free to use. This bridge lets you tap into those credits from Claude Code.

License: MIT Node.js Platform


Why?

Devin by Cognition provides access to powerful models like GLM-5.2 High, but its API isn't directly compatible with Claude Code. GLM-5.2 is included as a promo free model with Devin Pro (Windsurf) subscriptions, making it essentially free to use. This bridge solves four incompatibilities:

Problem Solution
GLM-5.2 errors on "You are Claude Code" identity Rewrites system prompt to generic assistant
GLM-5.2 has no native tool_use support Converts tools ↔ text instructions (XML tags)
Cognition API content filter blocks Claude Code prompts Sanitizes security/billing/imperative language
GLM-5.2 can't see images Routes images to kimi-k2-7/swe-1-7 (via ACP+PIL) for description, then GLM-5.2 answers using the description

Architecture

                         Claude Code (Anthropic API)
                                    ↓
                         glm-proxy (port 20130)
                      ← Rewrites prompts, converts tools,
                         detects images in requests
                                    ↓
              ┌─────────────────────┴──────────────────────┐
              ↓                                            ↓
     Text-only requests                           Image requests
              ↓                                            ↓
     9router (port 20128)                    Phase 1: Describe image
              ↓                            → windsurf-server (port 8083)
     windsurf-server                           kimi-k2-7 = "eyes"
     (port 8083)                               (via ACP + PIL/ImageMagick)
              ↓                                            ↓
     GLM-5.2 answers                       Phase 2: Replace image with
     (reasoning + tools)                    text description, then forward
                                            ↓
                                   9router (port 20128)
                                            ↓
                                   windsurf-server (port 8083)
                                            ↓
                                   GLM-5.2 = "brain"
                                   (answers using description)

Vision: "Eyes + Brain" pattern

GLM-5.2 doesn't support vision. When Claude Code sends an image, the proxy uses a two-phase approach:

Phase 1 — "Eyes" (describe): Send image directly to windsurf-server:8083 → kimi-k2-7 analyzes pixel data via ACP path (PIL/ImageMagick) → returns a text description. Falls back to swe-1-7 if kimi fails.

Phase 2 — "Brain" (answer): Replace image blocks with [Image content]: <description> in the messages, then forward the text-only request through 9router → GLM-5.2, which answers normally using the description.

Both phases go through windsurf-server (port 8083), but Phase 1 goes direct (bypasses 9router, since 9router strips images), while Phase 2 goes through 9router (normal text path, preserves tools/system prompt/streaming).

Prerequisites

Tool Install Required?
Node.js 18+ brew install node ✅ Required
9router npm install -g 9router ✅ Required
Claude Code npm install -g @anthropic-ai/claude-code ✅ Required
Devin CLI See Devin docs ✅ Required
ClaudeKit Optional ⬜ Optional

Devin must be authenticated: Run devin auth login before setup.

Quick Start

# Clone
git clone https://github.com/mxuanvan02/devin-9router-bridge.git
cd devin-9router-bridge

# Start 9router (if not already running)
9router start

# Run setup — installs proxy, configures 9router + Claude Code
./scripts/setup.sh

# Restart Claude Code, then test
claude

In Claude Code, any prompt will now route through GLM-5.2 via Devin.

Configuration

Ports (configurable via environment variables)

Variable Default Description
GLM_PROXY_PORT 20130 glm-proxy listen port
ROUTER_PORT 20128 9router port
WINDSURF_PORT 8083 windsurf-server port
VISION_MODELS kimi-k2-7,swe-1-7 Vision models (comma-separated fallback)
VISION_HOST 127.0.0.1 windsurf-server host for vision
VISION_PORT 8083 windsurf-server port for vision
GLM_PROXY_PORT=20131 ROUTER_PORT=20129 ./scripts/setup.sh

Claude Code Settings

After setup, ~/.claude/settings.json is updated:

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:20130",
    "ANTHROPIC_API_KEY": "your-9router-api-key",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "ws/glm-5-2",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "ws/glm-5-2",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "ws/glm-5-2"
  },
  "model": "ws/glm-5-2"
}

Available Models

Model ID Backend Context Window Max Output Vision Promo Free?
glm-5-2 GLM-5.2 High (Cognition) 128K 200K Via kimi/swe ✅ Yes (tier 4)
swe-1-7 SWE-1.7 (Cognition) 128K 262K ✅ Native (ACP+PIL) ✅ Yes (tier 4)
swe-1-7-lightning SWE-1.7 Lightning (Cognition) 96K 202K ✅ Native (ACP+PIL) ✅ Yes (tier 2)
kimi-k2-7 Kimi K2.7 (Moonshot) 16K 262K ✅ Native (ACP+PIL) ✅ Yes (tier 4)

Note: "Context Window" is the input token limit. "Max Output" is the maximum output tokens. These are distinct limits — total tokens (input + output) can exceed the context window. Values are from the live Devin API (GetCliModelConfigs), not the model's native specs.

Vision = model can analyze images via ACP path (Devin CLI agent uses PIL/ImageMagick). GLM-5.2 itself can't see images — the proxy routes image requests to kimi-k2-7/swe-1-7 for description, then GLM-5.2 answers using the text description ("eyes + brain" pattern).

💡 All 4 models are promo free with Devin Pro (Windsurf) — they consume free credits, not paid quota. Paid variants like glm-5-2-1m (1M output), glm-5-2-max (max effort) require paid credits (tier 1). Default config uses glm-5-2 for all Claude Code tiers (Opus/Sonnet/Haiku).

Add more models by editing proxy/windsurf-server.js.

Project Structure

devin-9router-bridge/
├── proxy/
│   ├── glm-proxy.js              # Anthropic API ↔ GLM-5.2 bridge
│   ├── windsurf-server.js        # Devin → OpenAI-compatible API server
│   └── windsurf-provider.js      # Core: sends to server.codeium.com
├── proto/
│   └── windsurf.proto            # Protobuf schema (reverse-engineered)
├── scripts/
│   ├── setup.sh                  # One-command setup
│   ├── auto-start.sh             # macOS launchd auto-start
│   ├── auto-start-linux.sh       # Linux systemd auto-start
│   └── auto-start-windows.ps1    # Windows Task Scheduler auto-start
├── docs/
│   ├── GETTING_DEVIN_TOKEN.md    # How to get Devin session token
│   ├── TROUBLESHOOTING.md        # Common issues & fixes
│   ├── ARCHITECTURE.md           # How it works (detailed)
│   ├── HERMES_INTEGRATION.md     # Hermes Agent + GLM-5.2 setup
│   └── OPENCLAW_INTEGRATION.md   # OpenClaw + GLM-5.2 setup
├── package.json                  # npm dependencies (protobufjs)
├── LICENSE
└── README.md

Debug

# Run proxy with debug logging
GLM_PROXY_DEBUG=1 node proxy/glm-proxy.js 20130 20128

# Full debug (logs system prompts + messages)
GLM_PROXY_DEBUG=2 node proxy/glm-proxy.js 20130 20128

# Check logs
tail -f /tmp/glm-proxy.log
tail -f /tmp/windsurf-server.log

Context window configuration

GLM-5.2 is available as a promo free model with Devin Pro (Windsurf) — you get free credits that cover GLM-5.2 usage. The free variant (glm-5-2) has a 128K context window (input) and 200K max output tokens. The glm-5-2-1m variant supports 1M max output tokens but requires paid credits (tier 1, costMultiplier 3x). Context window remains 128K on all variants. The proxy defaults to no truncation (full context passthrough).

If you hit content filter or quota limits on the promo tier, set truncation limits via env vars:

# Free tier: truncate system prompt to 1500 chars, messages to 3000 chars
GLM_PROXY_MAX_SYSTEM_LEN=1500 GLM_PROXY_MAX_MSG_LEN=3000 node proxy/glm-proxy.js

# Unlimited (default): no truncation
node proxy/glm-proxy.js
Env var Default Description
GLM_PROXY_MAX_SYSTEM_LEN 0 (no limit) Max chars for system prompt (0 = no truncation)
GLM_PROXY_MAX_MSG_LEN 0 (no limit) Max chars per message (0 = no truncation)
GLM_PROXY_DEBUG 0 Debug logging level (1=basic, 2=verbose)

Auto-start (all platforms)

OS Script Service type
macOS ./scripts/auto-start.sh launchd plist
Linux ./scripts/auto-start-linux.sh systemd user service
Windows powershell -ExecutionPolicy Bypass -File scripts/auto-start-windows.ps1 Task Scheduler

Each script installs the glm-proxy as a background service that starts at login and auto-restarts on failure.

Using with Hermes Agent

Hermes Agent can use GLM-5.2 through glm-proxy for tool calling (which GLM-5.2 does not support natively in OpenAI format). See docs/HERMES_INTEGRATION.md for full setup guide.

Quick config — add to ~/.hermes/config.yaml (or %USERPROFILE%\.hermes\config.yaml on Windows):

model:
  default: ws/glm-5-2
  provider: nine-glm
  api_key: YOUR_9ROUTER_API_KEY
  base_url: http://localhost:20130    # glm-proxy, NOT 20128
  api_mode: anthropic_messages        # glm-proxy speaks Anthropic format
  context_length: 128000
  max_tokens: 200000

Using with OpenClaw

OpenClaw can use GLM-5.2 through glm-proxy for tool calling (which GLM-5.2 does not support natively in OpenAI format). See docs/OPENCLAW_INTEGRATION.md for full setup guide.

Quick config — add a glm-proxy provider to openclaw.json:

{
  "models": {
    "providers": {
      "glm-proxy": {
        "baseUrl": "http://localhost:20130",
        "api": "anthropic-messages",
        "apiKey": "YOUR_9ROUTER_API_KEY",
        "auth": "api-key",
        "models": [
          {
            "id": "ws/glm-5-2",
            "api": "anthropic-messages",
            "contextWindow": 128000,
            "maxTokens": 200000
          }
        ]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "glm-proxy/ws/glm-5-2"
      }
    }
  }
}

An automation script (fix-openclaw-glm-proxy.py) is included in the guide to migrate existing GLM models from 9router to glm-proxy automatically.

Limitations

  • Token expiry: Devin CLI refreshes tokens in memory but doesn't write them back to credentials.toml. If you get "Cascade session" errors, run devin auth login again and restart windsurf-server.
  • Streaming-focused: Non-streaming responses are not fully tested.
  • Content filter: The Cognition API content filter is aggressive — some complex prompts may still be blocked.
  • Single tool per turn: GLM-5.2 may not reliably handle multiple tool calls in one response.
  • Vision (OCR): kimi-k2-7/swe-1-7 analyze images via PIL/ImageMagick (pixel-level). They handle colors, shapes, dimensions, and simple charts well, but cannot OCR text in images (PIL doesn't include Tesseract). For OCR, install tesseract and use the ai-multimodal skill.
  • Vision (complex scenes): Complex images (detailed scenes, multi-object photos) may take 2-3 minutes per image due to ACP agent's pixel-by-pixel analysis. Simple images (solid colors, shapes) take ~10-15s.

Uninstall

# macOS
launchctl unload ~/Library/LaunchAgents/com.devin-9router-bridge.glm-proxy.plist 2>/dev/null
rm ~/Library/LaunchAgents/com.devin-9router-bridge.glm-proxy.plist

# Linux
systemctl --user stop devin-9router-bridge-glm-proxy 2>/dev/null
systemctl --user disable devin-9router-bridge-glm-proxy 2>/dev/null
rm ~/.config/systemd/user/devin-9router-bridge-glm-proxy.service

# Windows (PowerShell)
Unregister-ScheduledTask -TaskName "Devin9RouterBridge-glm-proxy" -Confirm:$false

# All platforms — kill processes and remove files
pkill -f "glm-proxy.js"
pkill -f "windsurf-server.js"
rm -rf ~/.devin-9router-bridge

# Restore Claude Code settings (backup was created during setup)
# Look for ~/.claude/settings.json.bak.* and restore the latest one

Contributing

Contributions welcome! See CONTRIBUTING.md.

License

MIT © 2026

Acknowledgments

About

Connect Devin models (GLM-5.2, etc.) to Claude Code + ClaudeKit via 9router. No Anthropic API key needed.

Resources

Code of conduct

Contributing

Stars

4 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages