agentconnect

AgentConnect is an open-source, multi-agent collaboration platform positioned as an alternative to Claude Tag. Its premise is that individual AI agents are already capable, but making several agents work well with a team — and with each other — is the harder problem. AgentConnect turns the usual glue code (message channels, cron jobs, credential handling, context stitching) into a platform. Where agents live Agents are linked to bots and repositories in Slack, Telegram, Discord, Lark/Feishu, QQ, GitHub, GitLab, Gitea and Linear, plus webhooks. Work can start from a message, issue, pull request, webhook or schedule, and people follow along in the conversations where the work already happens. Runtimes and models Agents can run on Claude Code, Codex, Grok Build, DeepSeek, Pi, GitHub Copilot, Devin, Qwen Code, OpenCode and other ACP-compatible runtimes. Each agent's runtime, model, workspace, tools and machine are configured independently, so swapping one runtime does not require rebuilding the surrounding workflow. Jev (TypeSafe) is used for agent routing and model selection. Core capabilities described in the README - Roles and agent-to-agent calls: create agents with different roles and let them call one another while people observe in shared conversations. - Decisions: reusable rules powered by Jev decide when agents respond, route new conversations to specialist agents, and pick a runtime and model per session. - Memory and knowledge: each agent has its own memory and skills, and reviewed Knowledge can be published for all agents to find on demand. - Boundaries: visibility of agents and sessions, permitted repositories and tools, and which other agents may be called are all configurable. - Self-hosting: the Apache-2.0 stack can be run in your own environment, keeping agent execution and workspaces under your operation. Example use cases cited include triaging support conversations with specialist routing, cross-workspace support spanning Telegram and Slack, recurring operations triggered by schedule or webhook, keeping private forks current with upstream changes, and customized code review where reviewers, runtimes and models are selected per pull request. Getting started Docker Compose starts the Web console, Control Plane, Relay and PostgreSQL; the default stack listens only on 127.0.0.1 and uses local no-auth mode for evaluation. A Helm chart is published per release at oci://ghcr.io/agentconnect-md/charts/agentconnect for Kubernetes installs. A Setup Server on loopback configures browser authentication (Logto), provider apps and sign-in methods, while deployment topology stays in compose.env or Helm values. The repository also ships a setup skill under .claude/skills/agentconnect-setup so coding agents can walk through configuration interactively. Architecture The Daemon runs placed agents over daemon-owned ACP, owns workspaces and session state, maintains direct platform connections and schedules, and sends model-provider traffic directly. The optional Relay accepts callback-based ingress and webchat, proxies centrally managed MCP and OpenConnector access, and forwards message ingress to the owning daemon without durable storage. The Control Plane and Web UI manage authentication, configuration, placement, permissions, metadata and observability, storing explicitly approved organization knowledge and skill revisions while otherwise proxying bounded daemon reads. Live platform messages and ACP update streams stay on the daemon/relay data plane; the Control Plane stores coordination metadata rather than message bodies or session streams, and established sessions continue if it is temporarily unavailable. Development requires Node >= 24.12.0 and pnpm 11, with pnpm install && pnpm dev running all packages in watch mode. The project is licensed under Apache 2.0 and offers a Slack community, documentation site, cloud option and blog.