Teams need agent coordination infrastructure, not another way to run a single agent

The core problem AgentConnect addresses is not "how do I run an AI agent" but "how does a team share one." Right now, most agents live in a single developer's terminal. Nobody else on the team can see what the agent is doing, take over a session, or build on the context it accumulated. So every team ends up writing the same scaffolding: message routing, credential handling, cron triggers, context stitching across tools. AgentConnect replaces that scaffolding with a platform.

The architecture is worth understanding before the feature list. A Daemon runs the actual agents, owns their workspaces and session state, and sends model traffic directly to providers—nothing routes through a central server. A Relay handles callback ingress and proxies managed tool access. The Control Plane manages auth, configuration, placement, and permissions, but deliberately stores coordination metadata rather than message bodies or session streams. If the Control Plane goes offline, established sessions and daemon-local schedules keep running. That's a meaningful design choice: the coordination layer is not the execution layer.

The agent model is role-based and runtime-agnostic. You configure each agent with its own model, instructions, tool access, and sandbox policy, then assign it to channels or repositories. Agents can call one another. A pull request can trigger a general code reviewer, which then pulls in a security reviewer only when the diff warrants it—each running a different model with different repository permissions. The platform speaks ACP (Agent Communication Protocol), so Claude Code, Codex, Grok Build, DeepSeek, and Pi can run side by side without rebuilding the workflow when you swap one out.

The channel integrations cover Slack, Telegram, Discord, Lark, GitHub, and GitLab. Work can start from a message, an issue, a pull request, a webhook, or a schedule. The README describes a cross-workspace support scenario where a conversation begins in Telegram, pulls in engineering from a trusted Slack workspace, and delivers the resolution back where it started—which illustrates the actual coordination problem this is solving rather than just listing integration logos.

Setup is Docker Compose for local evaluation, Helm for Kubernetes. The default stack listens only on loopback with no-auth mode, which is sensible for first contact. The repo also ships a setup skill at .claude/skills/agentconnect-setup that coding agents like Claude Code can pick up directly—an interactive tutorial that verifies each checkpoint and never asks you to paste secrets into chat.

Development requires Node 24.12.0 or higher and pnpm 11. The project is Apache 2.0 licensed and self-hostable, which matters if you care about where agent execution and workspace data actually live.

The honest caveat: this is infrastructure for teams that already have agents doing real work and need to share them. If you're still evaluating whether agents are useful at all, the coordination layer is premature. But for teams past that threshold, the separation of execution from coordination—and the explicit design choice to keep the Control Plane out of the hot path—is the kind of architectural thinking that tends to age well.