AgentGate

One local gateway for your AI model requests.
AgentGate is a local AI gateway for AI apps and clients, including Codex, Claude Code, Gemini CLI, OpenCode, AtomCode, and apps compatible with OpenAI, Anthropic, or Gemini protocols. It routes each request to the model you choose, fails over automatically when a provider breaks, traces everything locally, bridges protocol differences across OpenAI-compatible providers, and keeps long-running AI tasks from being interrupted by automatic sleep on macOS and Windows.

中文 · Download · 5-Minute Quick Start · Full Reference · 💬 Discussions

GitHub: dengmengmian/agentgate-ai

New in v1.6.0 — Session affinity, budget gate, local models: Stick multi-turn chats to one upstream, cap daily spend (notify / block / force cheapest), Diff raw vs converted request bodies with redacted repro export, and one-click scan for Ollama / LM Studio. See the v1.6.0 release notes · Cursor / Continue / Cline · Local models.

Download

Your machine	Download
macOS Apple Silicon	AgentGate_1.6.1_aarch64.dmg
macOS Intel	AgentGate_1.6.1_x64.dmg
Windows 10 / 11	AgentGate_1.6.1_x64-setup.exe
Debian / Ubuntu	AgentGate_1.6.1_amd64.deb
Other Linux distros	AgentGate_1.6.1_amd64.AppImage

Windows install note: Edge/Chrome may say the file is “usually not downloaded,” and SmartScreen may show Windows protected your PC. That is expected: the installer is not Authenticode-signed. There is no free, widely trusted Windows code-signing certificate (Let’s Encrypt only covers HTTPS, not .exe), so open-source builds often ship unsigned — this is not a virus report. Download only from GitHub Releases. In the browser choose Keep / Keep anyway, then when SmartScreen appears click More info → Run anyway.

On macOS you can also install with Homebrew:

brew install --cask dengmengmian/tap/agentgate

Headless CLI (agentgate-serve) tarballs and older versions are on the Releases page.

Why AgentGate

Official experience intact	Model routing is yours	Every request visible
Keeps Codex / Claude Code / Gemini CLI / OpenCode / AtomCode usable the way you already use them, with one-click restore to official configs.	Official client requests enter AgentGate first, then route through protocol conversion or native pass-through to the upstream provider you choose.	Route decisions, converted payloads, upstream errors, tokens, cost, latency, and failover attempts are traced locally.

AgentGate is not a hosted API reseller or a generic proxy. It is a local entry point for AI model requests — every request enters AgentGate first, then routes, converts, fails over, or passes through under your control.

Three focused tools, one workflow

CodeLeveler writes the code. ReviewGate reviews it. AgentGate connects both to your model APIs. Each tool works independently, or they can be used together:

Tool	Focus
AgentGate	Adapt model APIs behind one local gateway
CodeLeveler	Inspect, edit, run, and verify code in the terminal
ReviewGate	Review code changes and surface high-confidence issues

5-Minute Quick Start

1. Download and install AgentGate.
2. Open Quick Setup or Providers, then paste your provider API key.
3. Click Start Gateway on Overview or Gateway. The default endpoint is 127.0.0.1:9090.
4. On Clients, click Apply Config for Codex, Claude Code, OpenCode, Gemini CLI, or AtomCode.
5. Send a test message in the client. Use Switch to Official or history rollback whenever you want to restore the original config.

Provider presets fill common base URLs, protocols, model defaults, and capability matrices. Most users do not need to touch model mapping or advanced endpoint fields at first.

Common Uses

Goal	What AgentGate does
Use Codex with DeepSeek or MiMo	Lets Codex keep sending official Responses API traffic, then converts or passes it through to the selected upstream.
Keep Codex Desktop plugins working	Preserves the official OpenAI-authenticated provider path while routing model requests through AgentGate.
Use Claude Code with DeepSeek / MiMo / Copilot	Supports Anthropic-compatible pass-through, model-name mapping, and optional GitHub Copilot provider setup.
Avoid provider outages or quota stalls	Tries failover providers on configured status codes, keywords, timeouts, and cooldown state.
Keep long AI tasks running unattended	Prevents automatic sleep on macOS and Windows by default, with optional request-aware control and quick system-tray toggles.
Understand every request	Stores raw and converted payloads, route decisions, upstream errors, token counts, latency, and estimated cost.

Supported providers

Provider	Type	Native Protocols	Provider-Specific Handling
Xiaomi MiMo	mimo	Chat + Anthropic	Multi-turn reasoning_content round-trip, region-aware tp-* host auto-routing, temperature strip in thinking mode, tool_choice non-auto strip, omni web_search strip, web_search builtin gated by matrix, Web Search Plugin auto-degrade / retry
DeepSeek	deepseek	Chat + Anthropic	Image stripping with explicit notice, DeepSeek V4 thinking history reasoning backfill, schema cleaning, message reordering
Anthropic (Claude)	anthropic	Anthropic	tool_use/tool_result, input_schema, thinking budget, native cache_control
GitHub Copilot	copilot	Chat + Anthropic	GitHub token → Copilot bearer exchange, x-initiator billing classification, Claude model dash→dot normalization
OpenAI	openai	Chat + Responses	None (Responses passthrough or Chat conversion)
Google Gemini	google_gemini	Chat	None
Kimi / Moonshot	kimi	Chat + Anthropic	web_search → builtin_function/$web_search; K3 uses reasoning_effort:max (no K2 thinking param); coding models keep thinking on/off control
MiniMax	minimax	Chat	Strip reasoning_effort / response_format, <think> extraction
GLM (Zhipu)	glm	Chat	Generic
DashScope (Qwen)	dashscope	Chat	Generic
SiliconFlow	siliconflow	Chat	Generic
Volcengine (Doubao)	volcengine	Chat	Generic
Baichuan	baichuan	Chat	Generic
StepFun	stepfun	Chat	Generic
SenseNova	sensenova	Chat	Drops null strict / response_format / non-function tools, merges system messages
Yi (01.AI)	yi	Chat	Generic
ModelScope	modelscope	Chat	Generic
xAI (Grok)	xai	Chat	Generic
Mistral	mistral	Chat	Generic
Groq	groq	Chat	Generic
Together	together	Chat	Generic
Fireworks	fireworks	Chat	Generic
Cerebras	cerebras	Chat	Generic
Perplexity	perplexity	Chat	Generic
Cohere	cohere	Chat	Generic
OpenRouter	openrouter	Chat	None
Custom	custom_openai_compatible	Chat	None (set Base URL yourself)

Guides

• Codex Desktop plugins with third-party APIs
• Codex + DeepSeek
• Codex + Xiaomi MiMo
• Claude Code + DeepSeek
• Claude Code + GitHub Copilot
• Gemini CLI
• OpenCode
• Full reference

Screenshots

Dashboard	Providers

Routes	Logs

Notes

• GitHub Copilot provider support is optional. Using Copilot subscriptions outside official clients may carry account risk; read the dedicated guide before enabling it.
• The gateway endpoint is 127.0.0.1:9090; localhost:1420 is only the development UI port.
• AgentGate is local-first and single-user. If you operate a shared API server or billing platform, tools like one-api, new-api, or LiteLLM may fit better.

Development

pnpm install
pnpm tauri dev

Useful checks:

pnpm test
pnpm build
cd src-tauri && cargo test

Community

• Questions and setup help: Discussions
• Bugs and provider requests: Issues
• Contributing: CONTRIBUTING.md

License

MIT