← Back to blog

Introducing Barycent

Why we're building a shared center for open source agent collaboration — and why that means filling gaps, not building another gateway.

A barycenter is the point two orbiting bodies share. The Earth and the Moon both circle a point between them, and neither one is that point. That's the idea behind this organization's name, and it's the idea behind everything we're building.

The ecosystem already has a lot of pieces

Over the last year, the open agent ecosystem has grown fast:

  • The Agentic AI Foundation (AAIF) hosts MCP, A2A, goose, agentgateway, and seven other projects.
  • AGNTCY gives us SLIM messaging, the OASF schema, an agent directory, and — most relevant to us — AgentBridge, which already wraps Claude Code, Codex, Copilot, Cursor, goose, and OpenCode as A2A agents.
  • Project NANDA (MIT) is building an index for discovering agents across registries.

None of that needed us to build it again. So we didn't.

What's actually missing

What we found, after a few weeks of research across these ecosystems, is a gap in the collaboration layer:

  1. Nothing stops an agent from reviewing its own code.
  2. Nothing caps spend before a 700-agent fan-out turns into a four-figure bill.
  3. Teams of agents from different vendors have no shared, enforced workflow — just side-by-side execution.

That's what Bary is for: a small CLI, built on AgentBridge, that runs a pipeline — plan, implement, cross-vendor review, test — and refuses to start if a review step would let an agent approve its own patch.

One organization, many working groups

Barycent isn't an "AI company" with a generic name bolted on. It's a general open source organization with working groups. The AI Working Group is active today. Nothing stops a Quantum Working Group, or something nobody has thought of yet, from forming the same way: a real gap, a working prototype, and maintainers willing to run it in the open.

Try it

cargo install --path crates/bary-cli
cd your-repo
bary init
bary run --task "add a login page" --mock

Everything is Apache-2.0, including the parts we haven't built yet. Read the design doc or follow along here on the blog.