Show HN: Async multi-agent coordination across physical environments

A developer released Agent Comms, an open-source protocol and toolkit that lets AI coding agents share context and coordinate tasks across sessions and machines, installable via pip from the GitHub repository BlahBlah23406/agent-comms. Agent Comms packages an agent's task roadmap, architectural decisions, rejected hypotheses, and uncommitted git diffs into portable "Context Capsules" for out-of-session handoff, and adds an AHRP Live Relay WebSocket mesh for peer discovery, publish/subscribe messaging, and cross-machine Remote Procedure Calls. The toolkit ships an MCP server exposing export_handoff_capsule, import_handoff_capsule, and list_saved_capsules, and works natively with Claude Desktop, Antigravity, and Cursor via the Model Context Protocol.

A lightweight protocol and toolkit for AI coding agents to share context, hand off tasks across sessions, and collaborate in real-time across machines. When working with AI coding agents Claude Desktop, Antigravity, Cursor, etc. , two major challenges arise: 1. Context Loss Across Sessions: Starting a new chat or moving to another machine resets the agent's mental model and working memory. 2. Multi-Agent Coordination: Agents running in different environments or physical computers cannot easily communicate, exchange state, or run coordinated tasks. Agent Comms provides a unified solution: - Context Capsules Out-of-Session Handoff : Packages an agent's task roadmap, architectural decisions, rejected hypotheses, and uncommitted git diffs into a compact, portable bundle. The next agent session resumes immediately without burning context tokens. - AHRP Live Relay In-Session Collaboration : A real-time WebSocket mesh supporting peer discovery, publish/subscribe messaging, and cross-machine Remote Procedure Calls RPC . - Native MCP Integration: Works out-of-the-box with Claude Desktop, Antigravity, and Cursor via the Model Context Protocol. Agent Comms captures cognitive state alongside your working tree without polluting Git commit history: - State & Decisions: Records what worked, what was rejected, and the immediate next steps. - Code Diffs: Captures staged, unstaged, and untracked changes into a clean patch. - Briefing Generation: Produces a token-efficient Markdown briefing tailored for the incoming agent. Machine A Active Session Machine B New Session ┌──────────────────────────┐ ┌──────────────────────────┐ │ Agent exports capsule │── File/Sync ──▶│ Agent imports capsule │ │ diffs + state + roadmap │ │ restores diffs + state │ └──────────────────────────┘ └──────────────────────────┘ For multi-agent workflows, a lightweight relay server coordinates agents over WebSockets: - Peer Discovery: Agents announce presence, roles, and hardware capabilities. - Cognitive Blackboard: Replicated state where agents share real-time decisions and learnings. - Direct RPC: Agents can invoke tools or run commands on peer machines. Using Pip: pip install git+https://github.com/BlahBlah23406/agent-comms.git Or Clone & Install Locally: git clone https://github.com/BlahBlah23406/agent-comms.git cd agent-comms pip install -e . One-Line Install Script: - macOS / Linux: curl -sSL https://raw.githubusercontent.com/BlahBlah23406/agent-comms/master/install.sh | bash - Windows PowerShell : irm https://raw.githubusercontent.com/BlahBlah23406/agent-comms/master/install.ps1 | iex Run initial setup: agent-comms setup Before ending a session or switching computers: agent-comms capsule pack \ --task "AUTH-01" \ --summary "Migrated auth module to JWT; integration test pending" \ --next "Run pytest tests/test auth.py" \ --learning "finding:PyJWT requires algorithms= 'HS256' " Restore your uncommitted files and task briefing: agent-comms capsule unpack "AUTH-01" See two local agents discover each other and collaborate: agent-comms demo Agent Comms includes an MCP server exposing export handoff capsule , import handoff capsule , and list saved capsules . Add to your MCP settings file: { "mcpServers": { "agent-comms": { "command": "agent-comms", "args": "mcp" } } } Once added, interact naturally with your agent: "Save my progress into a handoff capsule for task AUTH-01." "Resume task AUTH-01 from my latest capsule." - Natural Language User Guide USER GUIDE.md https://github.com/BlahBlah23406/agent-comms/blob/master/USER GUIDE.md — Plain-English prompt examples for Claude Desktop, Antigravity, and Cursor. - Onboarding Guide ONBOARDING.md https://github.com/BlahBlah23406/agent-comms/blob/master/ONBOARDING.md — Step-by-step developer onboarding and distributed agent recipes. - Protocol Specification SPECIFICATION.md https://github.com/BlahBlah23406/agent-comms/blob/master/SPECIFICATION.md — Formal AHRP wire protocol and JSON schemas. - Tutorial & Recipes TUTORIAL.md https://github.com/BlahBlah23406/agent-comms/blob/master/TUTORIAL.md — Hands-on walkthroughs and implementation examples.