An AI API proxy and routing gateway written in Go. It exposes OpenAI- and Anthropic-compatible endpoints to local CLI tools and routes requests to multiple upstream providers (OpenAI Codex, Google Antigravity, AWS IAM SSO / Builder ID, OpenAI-compatible gateways, AgentRouter, web-search services, and the Gommo AutoAI media API).
OmniProxy is derived from the SuperKiro project, extended with model-family grouped catalogs, bi-directional AgentRouter protocol transcoding, account-pool rotation with per-model failure isolation, and a local admin dashboard.
English | Tiếng Việt | 中文
- Protocol translation: Serves
/v1/chat/completions,/v1/messages(Anthropic),/v1/responses, and/v1/models. - Authentication methods:
- OpenAI Codex OAuth — browser PKCE flow, automatic token refresh, quota-window tracking.
- Google Antigravity OAuth — browser PKCE flow against Cloud Code Assist, per-account project discovery, and import of credentials an installed Antigravity / Gemini CLI already wrote locally. See the note in §7 before enabling it.
- AgentRouter — converts the agent payload format, maps
agent_thoughtstream events toreasoning_content, and maintainsX-Agent-Session-IDacross turns. - OpenAI-compatible gateways — any external endpoint, with model catalog discovery from
/v1/models, optional per-account Responses-API dialect (externalApiDialect), plan-window quota bars for new-api gateways (5h/7d subscription usage), and undeclared request parameters forwarded verbatim so nothing the IR cannot model is silently dropped. - AWS IAM SSO / Builder ID — login and background token refresh for CodeWhisperer/Kiro.
- Service API keys — web search via Firecrawl, Tavily, Exa, Jina Reader.
- Gommo AutoAI — long-lived token for the media API behind
api.gommo.net(also served by the 79AI front end).
- Media generation: Image (
/v1/images/generations), speech (/v1/audio/speech), video (/v1/videos/generations, with/v1/videos/{id}for a render that outlives the request), and music (/v1/music/generations, with/v1/music/{id}for the status) are served from providers that expose them. Asynchronous upstream jobs are polled internally, so a client receives a finished result rather than a job id. - Model family catalog: Groups discovered model IDs into families (
gpt,claude,qwen,deepseek,glm,grok,llama,kimi,minimax) with context and output token limits. - Account pool:
- Selection strategies: weighted round-robin, cost-optimized, reset-aware.
- Per-model cooldown, so a rate-limited model does not disqualify the whole account.
- Sticky routing per conversation to reuse the upstream prompt cache.
- Local dashboard: Embedded web UI (EN / VI / ZH), daily token and credit usage, API key management, fallback combos.
[Clients: Claude CLI, Codex CLI, any OpenAI-compatible client]
│
▼ (/v1/chat/completions, /v1/messages, /v1/models)
┌────────────────────────────────────────────────────────────────────────┐
│ OmniProxy │
│ │
│ ┌──────────────────────┐ ┌─────────────────┐ ┌──────────────────┐ │
│ │ Inbound Handler │ │ Model Catalog │ │ API Key & Quota │ │
│ │ (OpenAI / Anthropic) │ │ (Family Groups)│ │ Manager │ │
│ └──────────┬───────────┘ └────────┬────────┘ └─────────┬────────┘ │
│ │ │ │ │
│ ▼ ▼ ▼ │
│ ┌──────────────────────────────────────────────────────────────────┐ │
│ │ Account Routing Pool │ │
│ │ - Round-Robin / Cost-Optimized / Reset-Aware │ │
│ │ - Per-Model Cooldown & Error Isolation │ │
│ │ - Sticky Prompt Cache Router │ │
│ └──────────────────────────────────┬───────────────────────────────┘ │
│ │ │
│ ┌──────────────┬───────────┼───────────┬──────────────┐ │
│ ▼ ▼ ▼ ▼ ▼ │
│ ┌───────────┐ ┌────────────┐ ┌───────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Codex │ │Antigravity │ │AgentRouter│ │ External │ │ Gommo │ │
│ │ OAuth │ │ (Gemini- │ │ (Protocol │ │ Adapters │ │ (Async │ │
│ │ (PKCE) │ │ shaped) │ │ Transcode)│ │ (OpenAI) │ │ media) │ │
│ └─────┬─────┘ └─────┬──────┘ └─────┬─────┘ └────┬─────┘ └────┬─────┘ │
└────────┼─────────────┼──────────────┼────────────┼────────────┼────────┘
│ │ │ │ │
▼ ▼ ▼ ▼ ▼
[OpenAI Codex] [Cloud Code [AgentRouter] [External [api.gommo.net]
Assist] Providers]
- Go 1.25+ (or Docker).
git clone https://github.com/mxuanvan02/OmniProxy.git
cd OmniProxy
go build -o omniproxy .
./omniproxyThe server listens on http://127.0.0.1:8080 (or the port in data/config.json). The dashboard is at http://127.0.0.1:8080/admin.
docker compose up -dPrebuilt binaries for Linux/macOS (amd64, arm64) and Windows (amd64) are attached to every GitHub Release. scripts/update.sh is the fast path — it checks the latest release, verifies the sha256, backs the current binary up under .rollback/, swaps in the new binary + web assets, restarts the service, health-checks it, and rolls back automatically if the new version does not come up:
./scripts/update.sh --check # what would change; nothing is touched
./scripts/update.sh # download → verify → backup → swap → restart → health-checkPoint it at your installation with OMNIPROXY_HOME (default ~/.omniproxy-user) and OMNIPROXY_PORT (default 8080). Restart goes through systemctl --user restart omniproxy-user.service; on macOS or custom setups set OMNIPROXY_RESTART_CMD instead. See the header of the script for every override.
A fresh install binds
127.0.0.1and generates a random admin password, printed once to stderr on first start — save it. Client requests are not key-checked by default: setrequireApiKey: trueand sethostto0.0.0.0only deliberately (the admin API exposes every stored upstream token).
{
"host": "127.0.0.1",
"port": 8080,
"password": "your-admin-password",
"requireApiKey": false,
"accounts": [
{
"id": "00000000-0000-0000-0000-000000000001",
"nickname": "AgentRouter Primary",
"authMethod": "agentrouter",
"provider": "AgentRouter",
"baseUrl": "https://agentrouter.org",
"accessToken": "sk-ar-...",
"enabled": true,
"weight": 1,
"region": "external"
},
{
"id": "00000000-0000-0000-0000-000000000002",
"nickname": "External Provider",
"authMethod": "external_openai",
"provider": "External OpenAI",
"baseUrl": "https://api.example.com",
"accessToken": "sk-...",
"enabled": true,
"weight": 1,
"region": "external"
},
{
"id": "00000000-0000-0000-0000-000000000003",
"nickname": "Gommo Media",
"authMethod": "gommo",
"provider": "Gommo AutoAI",
"accessToken": "<gommo-access-token>",
"gommoDomain": "79ai.net",
"gommoProjectId": "default",
"imageModel": "<image-model-id>",
"gommoTtsModel": "eleven_flash_v2_5",
"gommoVoiceId": "<voice-id>",
"capabilities": ["image", "video", "audio-tts", "audio-music"],
"enabled": true,
"weight": 1,
"region": "external"
}
]
}gommoDomain is part of the Gommo credential rather than an optional setting: the upstream sends it as a body field on every call and rejects a request without it. A Gommo account must not carry the chat capability — it generates media and cannot answer a completion, so the chat pool would route a request it can never serve.
Antigravity accounts are written by the login/import flow rather than by hand; googleProjectId is discovered per account and should not be copied between accounts.
Accounts can also be added from the dashboard (/admin → Add Account), which handles the OAuth and import flows.
Point the Anthropic base URL at OmniProxy:
export ANTHROPIC_BASE_URL=http://127.0.0.1:8080
export ANTHROPIC_API_KEY=<your-omniproxy-api-key> # any value when requireApiKey is false
claude --model claude-opus-5 -p "Hello"Or persist it in ~/.claude/settings.json:
{
"env": {
"ANTHROPIC_BASE_URL": "http://127.0.0.1:8080"
}
}Add OmniProxy as an OpenAI-compatible provider in ~/.codex/config.toml:
model = "gpt-5.6-sol"
model_provider = "omniproxy"
[model_providers.omniproxy]
name = "OmniProxy"
base_url = "http://127.0.0.1:8080/v1"
env_key = "OMNIPROXY_API_KEY"export OMNIPROXY_API_KEY=<your-omniproxy-api-key>
codex exec "Hello"# Anthropic Messages
curl http://127.0.0.1:8080/v1/messages \
-H "Content-Type: application/json" \
-H "anthropic-version: 2023-06-01" \
-H "x-api-key: $OMNIPROXY_API_KEY" \
-d '{"model":"claude-opus-5","max_tokens":128,"messages":[{"role":"user","content":"Hello"}]}'
# OpenAI Chat Completions
curl http://127.0.0.1:8080/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $OMNIPROXY_API_KEY" \
-d '{"model":"gpt-5.6-sol","messages":[{"role":"user","content":"Hello"}]}'
# Available models
curl http://127.0.0.1:8080/v1/modelsThese routes need an account that advertises the matching capability (see the Gommo entry in §4).
# Image — answers with a URL per image
curl http://127.0.0.1:8080/v1/images/generations \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $OMNIPROXY_API_KEY" \
-d '{"prompt":"a paper boat on still water","size":"1024x1024","n":1}'
# Speech — answers with audio bytes, as the OpenAI route does
curl http://127.0.0.1:8080/v1/audio/speech \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $OMNIPROXY_API_KEY" \
-d '{"input":"Xin chào","voice":"<voice-id>"}' \
--output speech.mp3
# Video — the render is polled internally; `id` lets you retrieve a slow one later
curl http://127.0.0.1:8080/v1/videos/generations \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $OMNIPROXY_API_KEY" \
-d '{"prompt":"a paper boat drifting downstream","ratio":"16_9"}'
curl http://127.0.0.1:8080/v1/videos/<id> \
-H "Authorization: Bearer $OMNIPROXY_API_KEY"
# Music — the render is polled internally; `id` lets you retrieve a slow one later
curl http://127.0.0.1:8080/v1/music/generations \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $OMNIPROXY_API_KEY" \
-d '{"prompt":"upbeat lofi piano","title":"Midnight Drive","lyrics":"la la la"}'
curl http://127.0.0.1:8080/v1/music/<id> \
-H "Authorization: Bearer $OMNIPROXY_API_KEY"go test -count=1 ./...Google's Antigravity terms permit access only through Google's own client. An account used through this proxy can be disabled by Google at any time, and that has happened to people in practice. Treat it as an expected outcome of enabling the provider, not an edge case.
The OAuth client is not shipped here. Antigravity's desktop client ID and secret
belong to Google, so the login reads them from ANTIGRAVITY_CLIENT_ID and
ANTIGRAVITY_CLIENT_SECRET (or the antigravityClientId /
antigravityClientSecret settings) and fails with a clear message when neither
is set. Take the values from your own installed IDE. Importing credentials that
Antigravity or the Gemini CLI already wrote locally needs neither.
What OmniProxy does about it:
- It sends the protocol fields the Cloud Code Assist API requires — the OAuth
token, the
Client-Metadatadescriptor, and the account's own project — and nothing beyond them. - It does not rotate client fingerprints, invent platform or api-client values, or send a hardcoded project id belonging to someone else. Several third-party clients do the last one; requests then bill against a project the caller has no claim to, which is the pattern enforcement looks for.
- When Google does disable an account,
classifyAntigravityFailurerecognises the response and marks the accountBANNED, so the pool stops selecting it instead of retrying a dead credential on every request.
No configuration makes this compliant with those terms. Decide whether to use it with that in mind; the other providers here are unaffected either way.
Released under the MIT License. Derived from the SuperKiro project.