⚡ GLM-ZAI-2API

OpenAI-compatible proxy for Z.AI (GLM models) — single Go binary.

Features • Quick Start • Usage • Configuration • Troubleshooting • Contact

English | Русский

---

Features

• 🚀 Single binary — server, captcha solver, and token store in one Go binary; no pipes, no external services
• 🔓 OpenAI-compatible — /v1/chat/completions and /v1/models; drop-in for OpenCode, Claude Code, Roo Code, and any OpenAI-format client
• 🧩 Tool calling — full function calling support for OpenCode and other IDEs; works out of the box
• 🛡️ In-process captcha — Aliyun Captcha V3 solved directly in-process, replacing the old named-pipe captcha-server
• 📊 Embedded dashboard — live HTML status UI served at the root URL
• 🌐 Cross-platform — Windows, macOS, Linux
• 📦 Standalone — pure-Go SQLite driver (modernc.org/sqlite), zero CGO dependencies

💡 Works with IDE out of the box! Perfect for AI-assisted development.

---

Quick Start

1. Download

Download the latest release for your platform:

• Windows (64-bit)
• Linux (64-bit)
• macOS Intel
• macOS Apple Silicon

Or build from source:

git clone https://github.com/D3-vin/GLM-ZAI-2API.git
cd GLM-ZAI-2API
go build -trimpath -ldflags="-s -w" -o glm-zai-2api .

2. Collect device tokens (requires Chrome/Chromium)

The captcha solver needs a pool of device tokens:

go build -o token-collector ./token-collector/
./token-collector --count 750 --out tokens.sqlite

3. Configure

cp .env.example .env
# Edit .env — set ZAI_TOKEN for GLM-5 models (see Configuration)

4. Run

./glm-zai-2api

Dashboard: http://localhost:5082

5. Test

curl -X POST http://localhost:5082/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer d3vin" \
  -d '{"model":"glm-4.7","messages":[{"role":"user","content":"Hello!"}],"stream":false}'

---

Usage

Streaming

curl -N -X POST http://localhost:5082/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer d3vin" \
  -d '{"model":"glm-4.7","stream":true,"messages":[{"role":"user","content":"Say hi"}]}'

Connect from an IDE

Setting	Value
Base URL	http://localhost:5082/v1
API Key	d3vin (the AUTH_TOKEN)
Model	glm-4.7 (guest) or GLM-5.1 (with ZAI_TOKEN)

💡 OpenCode users: This API works perfectly with OpenCode! Just add it as a custom OpenAI endpoint and enjoy AI-assisted coding with GLM models.

Models

Model	Guest	With ZAI_TOKEN
glm-4.7	yes	yes
GLM-5-Turbo	no	yes
GLM-5v-Turbo	no	yes
GLM-5.1	no	yes
glm-5.2	no	sometimes

glm-4.7 works for guest accounts. All GLM-5 models require a real ZAI_TOKEN.

API endpoints

Method	Path	Auth	Description
POST	/v1/chat/completions	yes	Chat (stream / non-stream, tool calling)
GET	/v1/models	yes	List models
POST	/features	yes	Toggle webSearch, thinking, imageGen, persistHistory
POST	/prompt	yes	Legacy: single text prompt
GET	/status	no	Z.AI session status
GET	/admin/health	no	Health-check
GET	/admin/stats	no	Stats
POST	/admin/session/clear	yes	Clear all session history

Session headers:

Header	Purpose
X-Session-Id	Conversation ID (default default)
X-Fresh-Session	true — new chat_id, empty history

Extra request fields: webSearch / search, deepThink.

Tool Calling

Z.AI has no native function calling. The proxy implements an adapter:

• tools → rewritten as a "unit test" prompt → model emits JSON → parsed into tool_calls
• 5 parsing strategies: JSON code blocks, direct JSON, multi-line JSON, <<TOOL>> tags, natural language
• Compact tool signatures (~2-3k chars instead of ~60k)

Token collector

./token-collector --count 750 --out tokens.sqlite
./token-collector --headless=false      # visible browser

Collects up to 1250 device tokens via headless Chrome (chromedp). Tokens are consumed by the captcha solver and removed after use.

---

Project Structure

├── main.go              # HTTP server, config, OpenAI routes, .env loader
├── zai.go               # Z.AI session, HMAC-SHA256 signature, SSE streaming
├── captcha.go           # Aliyun Captcha V3 (in-process)
├── tokens.go            # SQLite device-token store
├── tools.go             # Tool calling adapter (unit test framing)
├── dashboard.go         # Embedded HTML dashboard
├── token-collector/     # Device token collector (chromedp → Chrome)
├── .env.example         # Configuration template
├── go.mod / go.sum
└── README.md

---

Configuration

Settings in .env (auto-loaded) or environment variables:

Variable	Default	Description
ZAI_TOKEN	(empty)	Z.AI JWT. Without it — guest mode (only glm-4.7)
PORT	5082	Server port
HOST	0.0.0.0	Bind address
AUTH_TOKEN	d3vin	Bearer token for protected endpoints
LOG_LEVEL	info	debug, info, warn, error
DB_PATH	tokens.sqlite	Device tokens database

Getting ZAI_TOKEN:

1. Open https://chat.z.ai and log in
2. DevTools (F12) → Application → Local Storage → token
3. Copy the value (starts with eyJ...)
4. Add to .env: ZAI_TOKEN=eyJ...

Optional CLI flags (override env / defaults):

• --db-path <path> — override DB_PATH (default tokens.sqlite)
• --verbose — enable captcha debug logging

./glm-zai-2api --db-path /custom/tokens.sqlite --verbose

---

Troubleshooting

Symptom	Cause	Fix
captcha verification failed — check tokens.sqlite	Token DB missing or empty	Run token-collector to rebuild tokens.sqlite
No device tokens available	All tokens consumed	Re-run the collector; each token is single-use
401 from Z.AI	JWT expired	Set a fresh ZAI_TOKEN, or restart (guest flow re-initializes)
Empty stream response	Captcha/network error	Set LOG_LEVEL=debug, inspect SSE logs
Model not available for current user level	Guest account + GLM-5 model	Set ZAI_TOKEN with a real account

---

Contact

• GitHub: https://github.com/D3-vin/GLM-ZAI-2API
• Telegram: @D3_vin
• Author: @D3vin_dev

---

License

Provided as-is for educational and interoperability purposes. Use responsibly and in accordance with Z.AI's terms of service.