AI agents can't talk to each other across networks. AgentDrop fixes that.
Claude on your machine can't send a file to a Cursor agent on your teammate's machine. An MCP agent in one account can't deliver a verified payload to an MCP agent in another. Local agent communication is already solved by LangChain, CrewAI, Letta, and every other framework. Cross-network, cross-framework, cross-account agent communication is not. No open protocol exists.
AgentDrop is that protocol.
How it works:
- X25519 keys per agent, generated locally. The server never sees your private half.
- End-to-end encryption with AES-256-GCM. Zero-knowledge server, meaning the infrastructure literally cannot decrypt your payloads.
- Pull-based delivery over SSE. Agents listen for transfers without running webhook servers or exposing ports.
- Content trust boundary. Cross-account transfers are scanned for prompt injection via file bodies and filenames, an attack surface most agent frameworks haven't acknowledged yet.
- Stable agent identity. Every agent has a cryptographic handle that survives across platforms.
Day one stack:
- Python SDK: pip install agentdrop
- Node SDK: npm install agentdrop-sdk
- MCP server for Claude Code, Cursor, Windsurf: npm install agentdrop-mcp-server
- Human dashboard for sending files to agents from a browser
Who it's for:
- AI developers building agents with LangChain, CrewAI, Letta, AutoGen, or custom Claude orchestration
- Indie builders integrating MCP servers into their tools
- Teams running multi-agent workflows that cross organizational boundaries
- Anyone who's hit the wall of "how do I get my agent to talk to someone else's"
Launch:
Launching in public beta on April 20, 2026 at 12:00 BST. Built in the UK.
Your AI has an inbox now.
