CrossLink

July 14, 2026 · View on GitHub

CrossLink Banner
Go Version Go Report Card License PRs Welcome
Stars Issues Discussions Last Commit

CrossLink

One Gateway. Every Model. Zero Lock-in.

OpenAI & Anthropic Compatible LLM API Gateway

English | 中文 | العربية

A unified proxy with true bidirectional protocol translation, error-classified failover, an MCP gateway with per-tool RBAC, pluggable guardrails, and a built-in admin dashboard — for OpenAI, Anthropic, Azure, DeepSeek, Qwen, Ollama, and any OpenAI-compatible provider.

Get Started · Highlights · Features · Architecture · Docs · Contributing


Every LLM provider has a different API format, auth mechanism, and feature set. Adapting your code for each one is tedious, error-prone, and locks you in.

CrossLink is a universal adapter between your application and any LLM provider:

  • One endpoint — Your code talks to a single API in either OpenAI or Anthropic format
  • Any provider — Requests are routed to OpenAI, Anthropic, Azure, DeepSeek, Qwen, Ollama, or any OpenAI-compatible service
  • True bidirectional translation — Full streaming SSE conversion between OpenAI, Anthropic, and the OpenAI Responses API, including tool use and extended thinking
  • Resilient by classification — Failover isn't blind retry. Errors are classified as persistent (quota/billing — one strike, long cooldown) or transient (rate limit — threshold-based), so retries go where they can actually succeed
  • Observable routing — Every response carries x-crosslink-fallback-* headers, and a routing-stats API shows configured-vs-actual traffic distribution so you can see drift

Highlights

What makes CrossLink different — each backed by code, not marketing.

  • 🔁 Bidirectional streaming translation — OpenAI ↔ Anthropic ↔ Responses API, all three directions, via a real state machine (not request-level rewrites). Handles thinking blocks, partial-JSON tool args, and token counting mid-stream. → Architecture
  • 🛡️ Error-classified failover — DB-backed rule table distinguishes persistent vs. transient failures; half-open single-flight probing prevents stampedes when a flaky provider returns. → Routing & Failover
  • 🔌 MCP gateway with per-tool RBAC — Not a thin proxy. Transport abstraction, tool discovery with singleflight caching, encrypted credentials, and allow/deny lists scoped by key, team, or role. → MCP Gateway
  • 🚧 Guardrails as a plugin registry — Plug in any engine (regex, external API, future ML) via RegisterEngine. Actions: block / log / mask. Per-model config, fail-open or fail-closed. → Architecture
  • 📊 Routing transparencyx-crosslink-fallback-model / x-crosslink-fallback-count headers on every response, plus a routing-distribution API showing configured-vs-actual weight, deviation, error rate, and latency per provider.
  • ❤️‍🩹 Self-healing dispatch counters — Per-(provider, model) concurrency/RPM limits via Redis Lua with a TTL heartbeat; a crashed process can't leave a provider stuck "busy".
  • 🇨🇳 GM national crypto + air-gapped ready — SM2/SM3/SM4 mode (including HMAC-SM3 JWT signing) and a self-hosted slider CAPTCHA. No reCAPTCHA/hCaptcha dependency — deploys fully offline for 信创 compliance. → Deployment
  • 🎁 Generous open core — The Community edition (Apache 2.0) ships 39 actions including MCP, RBAC, routing stats, and error rules. Pro adds guardrails/playground/secrets; Enterprise adds multi-org, audit, and budgets.

Features

Core Gateway

  • Dual Protocol/v1/chat/completions (OpenAI) and /v1/messages (Anthropic) with automatic bidirectional translation, streaming included
  • Multi-Provider — OpenAI, Anthropic, Azure OpenAI, DeepSeek, Qwen, Moonshot, Ollama, and any OpenAI-compatible provider
  • Intelligent Routing — 6 strategies: weighted random, round-robin, least latency, least cost, least busy, and canary deployment
  • Automatic Failover — Multi-provider fallback chains with circuit breakers, configurable retry policies (exponential/fixed/linear backoff), and error classification
  • Response Caching — Redis-based caching with per-model TTL, gzip compression, and cache key isolation per user

Security & Control

  • Rate Limiting — Per-key RPM/TPM limits with global concurrency control (2000)
  • RBAC — Role-based access control for providers, models, API keys, and MCP
  • Budget Management — Per-key and per-team budget limits with automatic circuit breaking
  • Guardrails — Pluggable content-safety engine framework with configurable rules and actions
  • Crypto Flexibility — Standard (SHA-256/RSA/AES) or Chinese national cryptography (SM3/SM2/SM4)

Observability

  • Usage Analytics — Token usage, cost tracking, latency metrics, cache hit rates, and fallback/retry counts per request
  • Prometheus Metrics — Built-in metrics endpoint for monitoring
  • OpenTelemetry — Distributed tracing support
  • Structured Logging — JSON logging with request context

MCP Gateway

  • Model Context Protocol — HTTP/SSE transport, tool discovery with caching, health checks
  • Permission Management — Per-principal tool access control (allow/deny by key, team, or role)
  • Call Logging — Comprehensive tool call logging with monthly partitioning and auto-cleanup

Operations

  • Vue 3 Admin Dashboard — Built-in web UI for providers, models, keys, usage, and MCP management (CrossLink-UI-Standard)
  • Multi-Instance — Redis Pub/Sub for provider registry sync and distributed round-robin
  • Graceful Shutdown — 5-phase drain: in-flight SSE streams → HTTP shutdown → worker flush → background goroutine cancellation → DB cleanup
  • One-Command Deploy — Docker Compose spins up gateway, frontend, PostgreSQL, and Redis in one command

Architecture

CrossLink Architecture

See Architecture for the request flow, the streaming translator state machine, the fallback engine's timeout budgeting, and the 5-phase graceful shutdown.


Dashboard Preview

Admin dashboard overview

The admin dashboard: request volume, cost, token usage, latency, error rate, and model distribution at a glance.

MCP server management

MCP server management: registry with transport types (HTTP/SSE/stdio), health status, and per-server configuration.

Provider and model configuration

Provider & model configuration: weight, priority, pricing, and routing strategy per model.


Quick Start

Prerequisites

  • Go 1.22+ (building from source)
  • PostgreSQL 14+
  • Redis 7+

Frontend (CrossLink-UI-Standard) and backend are built together. One command starts everything:

git clone https://github.com/HotRiceNoodles/CrossLink.git
cd CrossLink
docker compose -f deployments/docker-compose.dev.yaml up --build

Frontend dashboard and API gateway are available at http://localhost (port 80).

China network? Use docker compose -f deployments/docker-compose.cn.yaml up --build with Go and npm mirrors pre-configured.

Build from Source

git clone https://github.com/HotRiceNoodles/CrossLink.git
cd CrossLink
cp configs/config.example.yaml configs/config.yaml
# Edit config.yaml with your database and Redis settings
make build
./bin/crosslink

Make Your First Request

Create an API key via the admin dashboard (http://localhost:8080), then try it in 30 seconds:

curl http://localhost:8080/v1/chat/completions \
  -H "Authorization: Bearer cl-your-api-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-chat",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

OpenAI SDK (Python)

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8080/v1",
    api_key="cl-your-api-key"
)

response = client.chat.completions.create(
    model="deepseek-chat",
    messages=[{"role": "user", "content": "Hello!"}]
)
print(response.choices[0].message.content)

Anthropic SDK (Python)

import anthropic

client = anthropic.Anthropic(
    base_url="http://localhost:8080",
    api_key="cl-your-api-key"
)

message = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello!"}]
)
print(message.content[0].text)

Configuration

All configuration lives in configs/config.yaml. Every value can be overridden with environment variables using the CL_ prefix (e.g., CL_DATABASE_HOST, CL_REDIS_PORT).

server:
  port: 8080
  read_timeout: 30s
  write_timeout: 120s

database:
  host: localhost
  port: 5432
  user: crosslink
  password: crosslink
  dbname: crosslink
  sslmode: disable

redis:
  host: localhost
  port: 6379

gateway:
  auth_key: "cl-change-me"

admin:
  username: admin
  password: changeme
  jwt_secret: "change-me-to-a-random-secret"

cache:
  enabled: true
  default_ttl: 5m

mcp:
  enabled: true
  health_check_interval: 30s

crypto:
  mode: standard    # standard (SHA-256/RSA/AES) or gm (SM3/SM2/SM4)

Provider Seeding

Providers are seeded from configs/providers.yaml on first run:

providers:
  - name: deepseek
    adapter_type: openai_compatible
    base_url: https://api.deepseek.com/v1
    api_key: ${DEEPSEEK_API_KEY}
    models:
      - name: deepseek-chat
        provider_model: deepseek-chat

API Endpoints

Gateway (API Key Required)

MethodPathDescription
POST/v1/chat/completionsOpenAI-compatible chat (stream & non-stream)
POST/v1/messagesAnthropic-compatible messages (stream & non-stream)
GET/v1/modelsList available models

MCP Gateway

MethodPathDescription
POST/mcp/:serverMCP JSON-RPC forwarding
GET/mcp/:serverMCP SSE transport

Admin (JWT Required)

MethodPathDescription
POST/admin/api/auth/loginLogin
CRUD/admin/api/providersProvider management (test via POST /:id/test)
CRUD/admin/api/modelsModel mapping management
CRUD/admin/api/keysAPI key management (regenerate via POST /:id/regenerate)
GET/admin/api/usageUsage logs with multi-dimensional filtering
GET/admin/api/routing/statsRouting distribution: configured vs. actual per provider
CRUD/admin/api/mcp/serversMCP server management
GET/admin/api/mcp/servers/:id/toolsList tools on MCP server

The full API reference — including request/response shapes, error codes, and the x-crosslink-fallback-* response headers — is in docs/api-reference.md.


Deployment

Production Docker Compose

docker compose -f deployments/docker-compose.prod.yaml up -d --build

China Network

Use the CN variant with Go proxy (goproxy.cn) and npm mirror (registry.npmmirror.com):

docker compose -f deployments/docker-compose.cn.yaml up --build

Nginx · Caddy · Systemd · GM

Production-ready Nginx config (TLS, security headers, SSE streaming) at deployments/nginx/, a Caddyfile for automatic HTTPS, a systemd unit, and a dedicated GM (SM2/SM3/SM4) deployment with GmSSL/Nginx + TLCP at deployments/gm/.

→ See Deployment for all options and multi-instance scaling notes.


Documentation


Roadmap

CrossLink is under active development. Current focus:

  • Provider guardrails with health-aware routing
  • Routing-distribution observability (/admin/api/routing/stats)
  • Self-healing concurrency counters (TTL heartbeat)
  • OpenAI Responses API translation (stream + non-stream)
  • Self-hosted slider CAPTCHA (air-gapped ready)
  • Provider guard alert rules (Enterprise) — in progress
  • Expanded guardrail engine ecosystem (ML-based classifiers)
  • Multi-team budgets and audit (Enterprise)

Have a request? Open a Discussion or a feature request.


Community & Support

  • 💬 Questions & ideasGitHub Discussions
  • 🐛 Bug reportsopen an issue (use the bug-report template)
  • 🔒 Security reports — see SECURITY.md for private disclosure
  • Like the project? Give it a star — it helps others find CrossLink.

Contributing

We welcome contributions of all sizes — bug fixes, features, docs, or ideas.

  1. Fork the repository
  2. Create a feature branch (git checkout -b feature/my-feature)
  3. Commit your changes (git commit -m 'Add some feature')
  4. Push to the branch (git push origin feature/my-feature)
  5. Open a Pull Request

See CONTRIBUTING.md for detailed guidelines.

Development

make build          # Build binary (bin/crosslink)
make run            # Run the server
make test           # Run all tests
make lint           # Run golangci-lint
make clean          # Remove build artifacts

Security

See SECURITY.md for our security policy and vulnerability reporting instructions.


License

Apache License 2.0