Centaur: Self-Hosted, Real-Time AI Agents for Teams

In an ecosystem where most AI agent platforms push toward SaaS and managed cloud, Centaur arrives with the opposite proposition: multiplayer agents that teams install and control on their own infrastructure. The initiative appeared on Hacker News on May 23 with minimal initial comment activity, but the concept touches a real pain point many engineering teams have been scratching for some time.

The premise is simple to state and complex to execute: why send sensitive context, proprietary code, or customer data to an external API when you can orchestrate the same agents from your own server?

What Centaur offers exactly

According to the project's own website, Centaur focuses on three pillars:

• Multiplayer: multiple users or processes can interact with the same agent or set of agents simultaneously, without state collisions.
• Self-hosted: the platform deploys on your own infrastructure. There's no dependency on a central service that could change pricing, terms of use, or simply disappear.
• Security: the execution model keeps context and credentials within the team's perimeter.

The project doesn't detail on its homepage which models it supports or whether there's native integration with the Claude ecosystem, MCP servers, hooks, or Claude Code subagents. However, the described architecture aligns well with the pattern Anthropic has normalized with MCP: a local server that exposes tools which the LLM calls on demand. Any platform that speaks MCP can theoretically connect to Centaur as an orchestration layer.

Why it matters now

Throughout 2025 and into 2026, the conversation about where agents live has gained weight. Claude Code, Anthropic's official CLI, has significantly democratized local agent execution with access to system tools. But Claude Code remains fundamentally an individual experience, or at most a single operator one. Coordination between multiple developers working on the same active agent isn't natively solved.

That opens a gap for tools like Centaur. Teams already using subagent and MCP server workflows need a coordination layer that doesn't force them through an intermediary SaaS. The self-hosted proposal also addresses growing concerns in regulated sectors, finance, healthcare, legal, where sending context to an external API can be directly infeasible due to security policy or compliance requirements.

Who it makes sense for

The most obvious profile is the engineering team already working with Claude-based agents or other LLMs that's hit the limit of single-user experience. It also fits companies with data residency requirements or those that simply don't want their conversation history with agents leaving their servers.

It's less relevant for small teams or individuals without their own infrastructure: the operational cost of maintaining a self-hosted server far exceeds the added value if there are no concurrent users or particularly sensitive data.

What we still don't know

The project is in very early stages, or at least that's how its public communication appears. The Centaur website doesn't publish extensive technical documentation, a list of compatible models, code license, or roadmap. Without those details, it's hard to assess whether this is a mature product, an MVP seeking traction, or a proof of concept not yet ready for production.

It also remains to be seen whether the team behind the project intends to formally integrate with Anthropic's MCP ecosystem or proposes its own abstraction layer. That architectural decision will significantly shape adoption among those already invested in existing MCP servers.

---

Our take: Centaur points to a legitimate need the ecosystem hasn't yet covered well. Whether it succeeds depends on whether the documentation and distribution model live up to the promise. For now, it's worth following closely before betting on it for production.