WorkBuddy Proxy

Local OpenAI-compatible API + web console for WorkBuddy desktop quota.

Chat completions and reasoning tokens come from WorkBuddy (POST https://www.workbuddy.ai/v2/chat/completions). This process only forwards: auth headers, account round-robin, optional HTTP egress, and reasoning_effort.

Requirements

• Node.js 18+
• A WorkBuddy account (login in the web UI, or import a desktop session file)

Quick start

Windows

copy config.example.json config.json
copy proxies.example.txt proxies.txt
npm install
node server.js

Or start.cmd.

Ubuntu / Debian

chmod +x start.sh install-ubuntu.sh
./start.sh

Foreground: ./start.sh
systemd user service:

./install-ubuntu.sh
journalctl --user -u workbuddy-proxy -f

Open http://127.0.0.1:8790/ → Accounts → Login browser.

Update (existing install)

git pull
npm install          # only needed when dependencies changed
# restart: start.sh / start.cmd, or: systemctl --user restart workbuddy-proxy

start.sh re-runs npm install automatically when node_modules is missing, so on Ubuntu a plain git pull && ./start.sh is enough.

The running version is shown in the console header (v1.1.0) and under Settings → Upstream. GET /health returns it as version.

OpenAI client

base_url = http://127.0.0.1:8790/v1
api_key  = value of proxyKey in config.json
model    = hy3 | hy4-preview | gpt-5.6-luna | …

If a model is set to high under Models, the proxy injects reasoning_effort even when the client omits it.

WorkBuddy chat is SSE-only. Non-stream OpenAI calls are still streamed upstream, then assembled.

Remaining credits

The console has a Credits tab showing how much quota is left, per package.

Under the hood it calls the same endpoint the WorkBuddy desktop app uses:

POST https://www.workbuddy.ai/v2/billing/meter/get-user-resource
{ "PageNumber": 1, "PageSize": 100, "ProductCode": "p_tcaca",
  "Status": [0, 3], "OnlyValidPeriod": true }

The response data.Response.Data.Accounts[] gives one entry per package with CycleCapacitySizePrecise (total), CycleCapacityRemainPrecise (left) and CycleEndTime (reset).

The proxy exposes it as:

Endpoint	Auth	Notes
GET /admin/credits	loopback only	used by the UI, ?force=1 bypasses cache
GET /v1/credits	Bearer <proxyKey>	?account=<email|uid>, ?force=1

Results are cached per account for 60 seconds. force=1 re-hits upstream.

curl -s -H "Authorization: Bearer $PROXY_KEY" \
  http://127.0.0.1:8790/v1/credits | jq '.accounts[0] | {usageLeft, usageUsed, usageTotal}'

Config (not committed)

Copy config.example.json → config.json.

Field	Meaning
proxyKey	Local API key for /v1
port / host	Default 127.0.0.1:8790
importAuthFile	Optional desktop workbuddy-desktop-ai.info path
effortByModel	Server-side reasoning per model id
egress	HTTP proxy pool (proxies.txt, host:port:user:pass)

Do not commit config.json, accounts.json, proxies.txt, or requests.jsonl.

Layout

server.js          HTTP server (OpenAI + admin + static UI)
egress.js          HTTP CONNECT pool (undici ProxyAgent)
public/index.html  console
start.sh           Ubuntu/macOS launcher
start.cmd          Windows launcher
install-ubuntu.sh  deps + systemd --user
CHANGELOG.md       release notes

See CHANGELOG.md for the version history.