Open source · MIT
One shell for every AI API
AI Switch brings the accounts, protocols, and capacity behind Codex, Claude Code, Gemini CLI, Grok and more into one place: a single local proxy entry, automatic pool scheduling, and on-demand upstream protocol bridging. The desktop app and the self-hosted web service share the same data.
- 7
- agent platforms
- 4
- upstream protocols
- 7
- bridge routes
- MIT
- licensed
Supported platforms
Codex, Claude Code, Gemini CLI and Grok get native config writing and official account import. OpenCode, OpenClaw and Hermes connect through generic API routing, which still covers routing, terminal launch and session workflows.
- CodexNative support
- Claude CodeNative support
- Gemini CLINative support
- GrokNative support
- OpenCodeGeneric API routing
- OpenClawGeneric API routing
- HermesGeneric API routing
Native config writing uses safe direct writes: a snapshot is taken before mutation, the write is atomic, concurrent changes are detected, and guarded rollback is supported.
See the full capability matrix →Protocol bridging, no CLI reconfiguration
Each CLI keeps speaking its own protocol. AI Switch translates the request inside the local proxy to whatever the upstream account actually supports, so switching providers never means editing CLI config.
Local entry
/responsesOpenAI Responses/v1/messagesAnthropic MessagesGemini nativeGemini accounts only127.0.0.1:19527Upstream protocol
openaiChat Completionsopenai-responsesResponses APIanthropicMessages APIgeminigenerateContentSeven bridge routes
Responses → Chat CompletionsResponses → ResponsesResponses → AnthropicResponses → GeminiClaude → Chat CompletionsClaude → ResponsesClaude → Gemini
Core capabilities
Pool scheduling
Give every account a priority from 1 to 5 and a concurrency cap, pooled per platform. High priority goes first; when it saturates, traffic falls through to the next tier.
Routing and bridging
Pick any of four upstream protocols. When the local entry protocol differs from the upstream, the proxy converts it. Listens on 127.0.0.1:19527 by default.
Failure handling
Transient failure counts and failure kinds are recorded, bad accounts are pulled out for a cooldown window, and retried back into the pool when it expires.
Usage visibility
Input, output and cache tokens are recorded per request, priced in both USD and CNY, and breakable down by account and platform.
Terminal and sessions
Launch CLIs straight from the built-in terminal, resume sessions, and swap between three Vibe skins when the mood calls for it.
Ecosystem
Manage MCP server config across 11 clients from one place, plus two bundled skill packages totalling 27 skills.
One UI, four shapes
Desktop and browser share the same React UI and the same database, so proxy behaviour is identical either way.
- 01
Desktop app
A native Tauri 2 shell talking to the Rust core over IPC.
- 02
LAN browser
Start the web service with token auth and open it from a phone or tablet.
- 03
Tailscale private
Log in to Tailscale manually and reach it across your private network.
- 04
Funnel public
Enable Funnel when you need public reach — the token is still required.
Default bind is 127.0.0.1:3090; binding to 0.0.0.0 must be explicit. Every /api/* and /ws/events request needs the access token, Tailscale included.
See deployment options →Architecture
One frontend, two transports, a shared core. The Rust core serves both desktop IPC and HTTP, so the two ends cannot drift apart.
Stack
- Tauri 2
- Rust
- axum
- sqlx
- SQLite
- React 18
- TypeScript
- Go sidecar
- Tailscale tsnet
- rustls
Get started with AI Switch
Install the desktop build and go, or run the service yourself for a team.