AI applications no longer rely on a single large language model. They orchestrate a growing mix of MCP-compatible agents, proprietary tools, and commercial LLMs, each optimized for a different task, latency profile, or price point. That additional power comes at a cost: more connections to manage, more auth flows to juggle, and more decisions about which model to call for every request.
Today we are launching Nexus, an open-source AI Router that eliminates that complexity. Nexus gives developers a single endpoint that:
- Aggregates every MCP server - stdio, SSE, or HTTP - behind one interface
- Routes each request to the best LLM for the job, balancing cost, latency, context length, and availability.
- MCP tool sprawl: LLMs tend to not find the right tool if there's too many tools available. Nexus solves this by indexing all available tools and providing a unified interface to search and execute the right tool.
- Security: Nexus secures your agents by enforcing policies and permissions, ensuring that only authorized users can access sensitive data or execute critical tasks.
- Cost control: Teams often discover avoidable LLM charges only after the bill arrives.
- Governance: Nexus provides a centralized platform for managing access, permissions, and policies, ensuring that your agents are used in a responsible and ethical manner.
- Observability: Nexus provides a unified interface for monitoring and debugging your agents, allowing you to track performance, identify bottlenecks, and troubleshoot issues.
| Capability | What it does | Why it matters |
|---|---|---|
| MCP aggregation | Nexus provides a search tool to find the right tool and execute it. | Your agents see one clean API instead of many bespoke ones. |
| LLM routing engine | It allows you to define the models and rate limits per provider, and provides a list of models from different providers from one OpenAI compatible API. | You control both performance and spend with policy, not hand-written if-else blocks. |
| Tool discovery | Exposes every downstream tool through the search and execute endpoints, namespaced by server. | Agents can self-discover capabilities across all your tools. |
| Security and governance | OAuth2, TLS, CORS, CSRF, and token forwarding are first-class. | Ship to production without bolting on a separate gateway or WAF. |
- Cleaner architecture: Run Grafbase for GraphQL Federation and schema composition, run Nexus for AI orchestration, and connect them with a single HTTP call.
- Faster iteration: Add new MCP servers or swap LLM providers without touching application code.
- Lower TCO: Keep inference bills predictable while safeguarding latency-sensitive paths.
- Enterprise-grade compliance: OAuth2-protected endpoints, TLS verification, and namespaced tool access align with enterprise security checklists.
- Adaptive routing algorithms that learn from historical cost and latency data
- Observability for throughput, token-spend, and success rates in the Grafbase Dashboard
- Custom policy DSL to express organization-specific routing logic
- Deeper Grafbase integration, letting organizations manage Nexus configurations and policies centrally.
Nexus is currently in beta. If your team is pushing the limits of multi-agent AI and wants early access to advanced routing policies, real-time analytics, or enterprise support, reach out. We are onboarding design-partners now and would love to learn about your use-case.
AI orchestration should feel invisible. With Nexus, it finally does.
Questions? Join our community on Discord, open an issue on GitHub, or Contact Us.