gemini-nexus/README.md

Gemini Nexus

Desktop gateway that pools Google accounts behind a local OpenAI/Anthropic-compatible API, with encrypted credentials and auto-switching.

Quickstart · Architecture · Security · Issues

_{English · Português · 中文 · Español}

---

📌 The problem

Working with AI-powered IDEs and coding tools on a single Google account means quota that runs out, manual account switching, no idea how many tokens you spent, and no view of what the proxy actually sends upstream. Gemini Nexus keeps all accounts in one pool, switches between them automatically and shows the traffic.

✨ What it does

• Pools Google accounts added through OAuth, with status (Active, Rate Limited, Expired), last-used time and an optional per-account proxy URL.
• Switches accounts automatically when quota drops below 5 % or an account is rate-limited; CloudMonitorService polls quota every 5 minutes.
• Schedules requests across the pool in cache-first, balance (default) or performance-first mode, with a per-account circuit breaker (default backoff 60 s → 5 min → 30 min → 2 h).
• Serves POST /v1/chat/completions (OpenAI), POST /v1/messages (Anthropic), GET /v1/models and GET /v1/models/capabilities, so Cursor, Windsurf, OpenCode, Claude Code and similar clients can connect.
• Maps requested model names to upstream models through the custom_mapping and anthropic_mapping settings; request timeout defaults to 120 s.
• Tracks token usage per account and model, charts prompt and completion tokens, and estimates cost with editable per-model pricing.
• Inspects traffic (latency, status, model), replays logged requests and generates cURL and Python snippets plus an IDE Quick Setup.
• Protects stored credentials with AES-256-GCM and an OS-keychain master key (see Security model).

🖼️ Demo

Screenshots

🚀 Quickstart

Use a release. Download the build for your platform from Releases, open the app, add at least one account, enable the proxy (it is disabled by default) and point your IDE at it:

API Base URL:  http://localhost:8045/v1   (default, configurable on the Proxy page; development builds use 8046)
API Key:       (copy it from the Proxy page in the app)
Model:         gemini-3-flash             (or any mapped model)

Warning

Builds are not code-signed. If Windows shows "Windows protected your PC", click More info → Run anyway.

Build from source.

Tool	Minimum version
Node.js	20+ (recommended in AGENTS.md; .nvmrc pins 22.17.1)
pnpm	10.11.0 (packageManager; lockfile is pnpm-lock.yaml)
Native toolchain	needed by better-sqlite3 and keytar: Visual Studio 2022 Build Tools (C++) on Windows, Xcode Command Line Tools on macOS, build-essential + python3 + libsecret-1-dev on Ubuntu (base-devel + python + libsecret on Arch)

git clone https://git.evandro.dev.br/evandro/gemini-nexus.git
cd gemini-nexus
pnpm install          # postinstall runs scripts/setup-dev.js --check-only to verify the toolchain
pnpm start            # dev mode (Electron + Vite HMR)
pnpm run make         # installers/packages for the current platform
pnpm run make:win     # Windows: extracts WiX from wix314-binaries.zip, then runs make

There is no .env.example and no variable is required. Optional variables read by the code or described in the docs:

Variable	Required	Description
HTTP_PROXY / HTTPS_PROXY / NO_PROXY	No	Route the app's outbound requests through a local proxy (see TLS inspection).
NODE_EXTRA_CA_CERTS	No	Extra CA certificate for Node-side TLS when inspecting traffic.
SENTRY_DSN	No	Sentry DSN used when error reporting is enabled.

🧱 Architecture

flowchart LR
    C["Client: Cursor, Windsurf, OpenCode, Claude Code"] -->|"/v1/chat/completions, /v1/messages"| P["NestJS proxy, port 8045"]
    P --> S["Account scheduler and circuit breaker"]
    S -->|"Google OAuth account"| G["Google upstream APIs"]
    M["CloudMonitor and AutoSwitch, quota every 5 min"] -->|"quota polling"| G
    S -->|"reads accounts"| DB[("SQLite: better-sqlite3 and Drizzle")]
    M --> DB
    K["AES-256-GCM, master key in OS keychain"] -->|"encrypts tokens"| DB

1. A client sends OpenAI- or Anthropic-format requests to the NestJS proxy (Fastify, bound to 0.0.0.0).
2. The proxy guard checks the API key when one is configured; with an empty key the proxy runs in open mode.
3. The token manager picks an account according to the scheduling mode and each account's circuit-breaker state, then calls Google's upstream APIs with that account's OAuth token.
4. CloudMonitorService polls quota every 5 minutes and AutoSwitchService switches the active account when quota is below 5 % or the account is rate-limited.
5. The React 19 renderer (TanStack Router and Query) talks to the main process over oRPC on IPC; accounts, usage and settings live in SQLite (better-sqlite3 + Drizzle). Design notes: docs/cloud_features.md.

🔒 Security model

Gemini Nexus holds OAuth tokens for several accounts and forwards traffic upstream, so credential handling is the core of the design. Everything below is implemented in this repository; limits are stated where they exist.

• Credential storage and isolation. token_json and quota_json are encrypted with AES-256-GCM (unique IV per operation, authenticated) before they reach SQLite. The master key lives in the OS credential store through keytar (Windows Credential Manager, macOS Keychain, Linux Secret Service/libsecret), never in a plain file. Legacy plaintext rows are migrated at startup. Accounts are kept in their own database (cloud_accounts.db), independent from the IDE's local database.
• Log masking. Keys such as token, refresh_token, api_key, authorization, cookie and client_secret are replaced by [REDACTED] before logging (src/utils/sensitiveDataMasking.ts); the traffic logger also redacts the authorization, x-api-key and cookie headers.
• Request logging, not an audit trail. Winston writes API, traffic and app logs to rotating files (API log: 30 days, 50 MB per file). These are operational logs, not a tamper-evident audit trail; the *-audit.json files are rotation bookkeeping from winston-daily-rotate-file.
• Proxy access. The proxy is off by default. Once enabled it listens on 0.0.0.0, and requests are authenticated only when api_key is non-empty (empty means open mode). Set a key and keep it private.
• TLS inspection (debugging only). docs/proxyman-install.md and docs/proxyman-debugging.md explain how to read the app's HTTPS traffic with Proxyman: trust its root CA, set the proxy variables and NODE_EXTRA_CA_CERTS, then start the app. NODE_TLS_REJECT_UNAUTHORIZED=0 (and the script's -AllowInsecureTls flag) is a temporary override; never leave it on. Do not commit proxy addresses, certificate paths, cookies or OAuth codes, and redact logs before sharing them. scripts/start-antigravity-with-proxyman.ps1 targets the installed official Antigravity client, not this app.
• Error reporting. Sentry reports (they need SENTRY_DSN) are gated by the error_reporting_enabled setting, which DEFAULT_APP_CONFIG ships as true; turn it off in Settings to opt out. The Windows user-name segment is masked in reported exception messages.

Supported versions and how to report a vulnerability: SECURITY.md.

🗂️ Structure

src/
├── components/   # UI (proxy/, usage/, ui/ primitives)
├── ipc/          # oRPC routers (account, cloud, config, gateway, usage...)
├── server/       # NestJS proxy gateway (modules: proxy, usage)
├── services/     # AutoSwitchService, CloudMonitorService, GoogleAPIService
├── routes/       # index, usage, proxy, settings
├── utils/        # paths, logger, sensitiveDataMasking
└── tests/        # unit tests (Vitest) and E2E (Playwright)
cli/              # separate Python CLI for account switching and quota checks
docs/             # design notes and Proxyman guides
scripts/          # setup-dev, build-windows, generate-assets

🧪 Scripts

Command	What it does
pnpm start	Run the app in dev mode (electron-forge start, with DEBUG=electron-forge:*).
pnpm run dev	Same as start.
pnpm run package	Package the app with Electron Forge, without building installers.
pnpm run make	Build installers/packages for the current platform.
pnpm run make:win	Windows build helper: extracts WiX from wix314-binaries.zip, adds it to PATH, runs electron-forge make.
pnpm run publish	Publish through Electron Forge.
pnpm run setup	Run scripts/setup-dev.js: detect the native toolchain and install or guide the setup.
pnpm run postinstall	scripts/setup-dev.js --check-only; runs automatically after pnpm install.
pnpm run lint	ESLint over the repo.
pnpm run format	Prettier check.
pnpm run format:write	Prettier write.
pnpm test	Vitest unit tests (vitest run).
pnpm run test:watch	Vitest in watch mode.
pnpm run test:unit	Vitest (default mode).
pnpm run test:e2e	Playwright E2E tests.
pnpm run test:all	Unit tests, then Playwright.
pnpm run type-check	tsc --noEmit.

📚 Documentation

• docs/cloud_features.md: account pool, quota monitoring and auto-switch design.
• docs/proxyman-install.md and docs/proxyman-debugging.md: TLS inspection setup (Windows).
• docs/OpencodeAPI.md: OpenCode custom-endpoint integration (in Portuguese).
• cli/README.md: the Python command-line tool.
• CONTRIBUTING.md, SECURITY.md, CODE_OF_CONDUCT.md, CHANGELOG.md, LICENSE, LICENSE-MIT.
• Translations: Português, 中文, Español.

🛣️ Limitations and next steps

• Upstream access goes through Google's v1internal Cloud Code endpoints and OAuth scopes (src/services/GoogleAPIService.ts). These are internal, undocumented APIs and may change without notice. The project is unofficial and not affiliated with Google or Anthropic.
• Accounts are added through Google OAuth; Claude models are reached through model mapping, not through separate Claude accounts.
• End-to-end tests are not run in CI (Electron needs a graphical environment); run pnpm run test:e2e locally.
• The Windows MSI is only produced when the WiX Toolset is available on the build machine; otherwise the Windows artifact is the ZIP.
• The workflows in .github/workflows/ are written for GitHub (they call the gh CLI and the GitHub release publisher), so this README shows no CI badge.
• Next step stated by the project: migrate the remaining original code to MIT (see License).

🤝 Contributing

Read CONTRIBUTING.md and CODE_OF_CONDUCT.md. Releases are versioned by semantic-release from Conventional Commits. Before opening a PR, these must pass:

pnpm run lint && pnpm run format:write && pnpm run type-check && pnpm test

📄 License

Mixed licensing: code inherited from AntigravityManager (by Draculabo) is CC BY-NC-SA 4.0 (non-commercial, share-alike); all new code is MIT. The project is migrating to MIT as the remaining original code is rewritten, so check both files before any commercial use.