AgentConnect: Open-Source Multi-Agent Platform Embeds AI Agents in Slack, GitHub, Feishu

Problem: Multi-Agent Collaboration in Existing Tools

As AI agents like Claude Code, Codex, and DeepSeek spread across teams, the challenge shifts from making agents smarter to making them collaborate like team members. Every team ends up building custom glue — message channels, cron jobs, credential management, context stitching, ACL visibility. AgentConnect aims to productize this glue layer.

Architecture: Control Plane / Data Plane Separation

The core design strictly separates:

Control Plane (CP) : Handles orchestration only — identity, routing, permissions, metadata, observability. It stays off the hot path of any real-time message.

Data Plane : Each daemon is a self-contained unit that processes platform messages (Slack, Feishu, etc.) and drives agents via the Agent Client Protocol (ACP). ACP never traverses the CP; local daemons use IPC (zero network hops), cloud pools use intra-cluster dialing.

This preserves the agent's native tool harness (terminal, file I/O, retrieval, MCP tools) as a single capability surface. The CP can fail while established sessions continue on the data plane (degraded availability), recovering later — a production-grade reliability pattern.

Empirical Evidence: LSP vs Grep Token Study

AgentConnect's lsp-vs-grep-token-study shows: when both grep and LSP (Language Server Protocol) are available, Claude models spontaneously choose LSP only 0–6% of the time; forcing semantic paths drops success from 100% to 89%. But with noisy codebases (e.g., hono), grep precision falls to 0.51 while LSP recovers ΔF1 +0.246 and saves 12% tokens. Conclusion: Agent capability = model × harness; ripping agents from their native tooling hurts precision and adds latency.

Trust Model: Three Orthogonal Levels

Operator-trusted Agent : Unsandboxed, explicitly defined as having the daemon OS account's environment permissions. Documentation warns against describing unsandboxed runtimes as restricted.

Per-agent sandbox : Opt-in isolation per agent at launch.

Force sandbox : daemon --require-sandbox enforces sandboxing for all agents with fail-closed behavior (daemon refuses work if sandbox fails).

Credentials are managed via the open-source OpenConnector gateway, keeping provider API keys out of agent processes.

Memory and Knowledge Management

Each agent has independent memory + skills, retaining cross-session learning.

Organizations can publish vetted Knowledge for any authorized agent to retrieve.

Backends: Native, Managed, Mem0 (external), or custom.

Critical boundary: agents only retrieve decisions from channels, repos, and systems they are authorized to access — no indiscriminate context flattening.

Relay: Optional Public Callback Entry

Stable public callbacks (Slack HTTP, Feishu HTTP, GitHub webhooks, webchat) route through a stateless Relay pool to the owning daemon, so daemons need not expose public ports. Direct bot connections (Slack Socket Mode, Feishu long-lived, Telegram) connect straight to daemons, bypassing Relay entirely. Relay is an optional ingress plane, not a mandatory hop.

Competitive Positioning

vs Claude Tag (cloud-hosted): AgentConnect is the open-source, self-hosted alternative — Apache-2.0, own infrastructure, any ACP agent, no vendor lock-in.

vs OpenClaw (personal terminal assistant): AgentConnect is team-grade — multi-user, multi-agent, shared sessions, per-member visibility, independent trust boundaries, agent-to-agent invocation.

vs Raft (standalone agent workspace): Raft requires migrating to a second IM; AgentConnect embeds agents into existing Slack/Feishu workflows.

vs Custom glue : Platformizes message channels, cron, credentials, context stitching, identity, routing, permissions, placement, triggers, delivery — so every team doesn't rewrite it.

Target Audience and When to Avoid

Suitable for: Engineering teams deeply collaborating in Slack/Feishu/Discord/GitHub and scaling multiple agents; teams currently writing custom agent glue; enterprises with data sovereignty/self-hosting/no-lock-in mandates; teams mixing vendors (Claude Code + Codex + DeepSeek + Grok Build) without rewriting workflows; mid-to-large teams demanding architectural rigor, observability, fine-grained permissions.

Avoid if: Solo developers wanting faster coding (OpenClaw or raw Claude Code lighter); teams wanting zero-ops hosted experience (self-hosted multi-component setup has learning curve); enterprises requiring battle-tested maturity (project ~7 weeks old, 1.3k stars, limited production validation — do POC first).

Risks and Current Status

Very young, high bus factor : Repo created 2026-07-24 (7 weeks ago), zfy0701 contributed 1100+ commits — core knowledge concentrated in few people.

Insufficient production validation : Architecture docs are production-grade but lack large-scale public deployments and benchmarks. Design maturity ≠ production proven.

Complexity barrier : Daemon, CP, Relay, sandbox, OpenConnector, Mem0 — Docker Compose starts simple; production (K8s + sandbox + HA) is complex.

ACP ecosystem dependency : "Any agent pluggable" relies on ACP protocol; gaps in runtime support reduce actual agent choice.

Low absolute stars : 1,351 stars vs mature competitors — limited third-party tutorials/integrations, self-service troubleshooting likely.

Recent velocity: v1.47 to v1.55 in two weeks (release every 2–4 days), commits focused on sandbox, security, Feishu, dependency hygiene — "growing skeleton while hardening for production" pace.

GitHub: github.com/agentconnect-md/agentconnect