A self-hosted gateway for the ZCode coding-plan API (GLM-5.3 / GLM-5.3-Flash), bundled with an account generator and a Cloudflare mail worker.
It exposes OpenAI- and Anthropic-compatible endpoints locally, pools multiple ZCode accounts behind one key, and can create new accounts for you on demand.
Unofficial, independent project. Not affiliated with ZCode or z.ai.
| Path | What it is |
|---|---|
zcode_gateway/ |
The gateway: FastAPI app, account pool, admin/dashboard API, Windows desktop shell, web dashboard |
zcode_gateway/dashboard/ |
The dashboard SPA (served at /dash) |
generator/ |
The account generator: signup, captcha solving, invite claim, proxy pool |
mail-worker/ |
Cloudflare Email Worker that receives signup verification mail |
tests/ |
Offline test suite (161 tests) |
tools/ |
import_account.py, smoke_live.py |
docs/ |
Protocol notes and diagnosis write-ups |
Double-click dist\ZcodeGateway.exe. On first launch it writes its own
config.yaml, accounts/, logs/ and gateway.db beside the executable
(portable — move the folder anywhere). Default port 8788.
Dashboard: http://127.0.0.1:8788/dash
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r requirements.txt
.\.venv\Scripts\python.exe -m zcode_gateway.mainPoint any OpenAI/Anthropic client at http://127.0.0.1:8788/v1 with the client
key from the Server tab.
{
"model": "GLM-5.3-Flash",
"messages": [{"role": "user", "content": "Hello"}],
"reasoning_effort": "high",
"stream": true
}Accounts → Login with Zcode (browser approval) or Import from Zcode app
(reads ~/.zcode/v2/credentials.json). The pool reloads automatically.
The Generator tab drives generator/ — it needs a real Python (not the
packaged exe) with the generator dependencies:
.\.venv\Scripts\python.exe -m pip install -r generator\requirements.txt
.\.venv\Scripts\python.exe -m playwright install chromiumThe gateway auto-detects the bundled generator/ folder. If you keep it
elsewhere, set it in Generator → Locate generator (applies live, no restart)
or generator.zcodeauto_dir in config.yaml.
Signups need verification mail. Use your own domain with the bundled worker:
- Add your domain to a free Cloudflare account.
- Generator → Email setup → step 2, click Deploy worker — it installs Node.js if missing, creates the KV namespace + secret and deploys.
- In Cloudflare: Email → Email Routing → Catch-all → Send to a Worker →
zai-mail. - Back in the dashboard: Test inbox.
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r requirements.txt -r requirements-dev.txt pyinstaller
.\.venv\Scripts\python.exe -m pytest tests -q
.\build-exe.ps1Output: dist\ZcodeGateway.exe (single file, no Python needed to run).
Live checks (explicit, separate from tests):
.\.venv\Scripts\python.exe tools\smoke_live.py
.\.venv\Scripts\python.exe tools\smoke_live.py --infer- OpenAI Chat Completions, legacy Completions and Models endpoints.
- Native Anthropic Messages, streaming and non-streaming.
- Thinking signatures, cache controls, tool blocks, images, system blocks and metadata preserved on native requests; full translation for the OpenAI adapter.
- Reasoning levels
low/high/max(medium→high,xhigh→max). - Per-model token quotas, reset times and plan expiry tracking.
405 / code 3012"unusual activity" is a per-model park: the account is skipped for that model only and the request rotates on. See docs/UNUSUAL-ACTIVITY.md.- Strategies: least-in-flight, drain-first, round-robin, LRU, quota-aware.
- Model overload (
529 / code 1305) retries before output with backoff + jitter. - Authenticated client/admin routes, per-install keys, redacted captures (off by default).
- One-click Cloudflare worker deploy with automatic Node.js install.
The generator's proxy tooling verifies every candidate against chat.z.ai
itself: the auth page must load with a valid certificate and the Aliyun
captcha host must be reachable. Around that gate:
- One parser for every list shape:
ip:port,scheme://,user:pass@, provider stylehost:port:user:pass, CIDR ranges (up to /16), HTML tables and JSON records (generator/proxy_parse.py). - TCP prefilter before any HTTPS request (measured: ~3 s/port instead of two full HTTP timeouts per dead host).
- Health scoring in
pool.json— a failed check adds one, a pass removes one; a proxy stays usable while its score is ≤ 3.proxies.txtis the best-first export, somain.pyis unchanged. Pasted proxies arevouchedand stay pending until a real check passes; only verified proxies are exported or counted by the generator. - Second chance pass over timeouts/resets (recovers roughly 30% more on free lists).
- Typed failure reasons (timeout, refused, TLS/MITM, hijacked page, wrong protocol, your PC ran out of connections, …) shown live in the dashboard instead of a bare "0 found".
- Dead cache (opt-in) so repeat scans skip hosts that just failed.
- Cross-process claims: eight workers are eight processes; the claim lock is a file lock, so two workers can never get the same exit IP.
- Per-source yield and scan presets (
balanced/thorough/fast). - Countries — the country of each proxy is looked up online (ip-api.com and
geojs.io in batches, with ipwho.is / ipinfo.io as fallbacks) and cached in
ip_country.json. Nothing is downloaded.
CLI: python generator/proxy_pool.py{ fetch | verify | recheck | pick | countries | stats | prune | clear-bad | clear-used | clear-dead }.
- Runtime secrets (
config.yaml,gateway.db,accounts/,generator/mail.json,*.har) are gitignored — they are never part of the repo or releases. - The gateway binds to
127.0.0.1by default and mints per-install random admin/client keys on first run. - Credential decryption uses the Windows AES-GCM implementation via CNG. On
non-Windows, encrypted imports require
cryptography.
AutoZCoder/
├─ zcode_gateway/ gateway package (+ dashboard/)
├─ generator/ account generator scripts
├─ mail-worker/ Cloudflare Email Worker template
├─ tests/ offline tests
├─ tools/ docs/ assets/
├─ config.example.yaml
└─ build-exe.ps1
MIT — see LICENSE.