- CollectionAgent Runtime & Infrastructure
- Classificationagent-native
- LicenseSDKs and protocol: Apache-2.0; daemon and relay: FSL-1.1-Apache-2.0 (per repo README license section)
- MCPAvailable
- Agent SkillUnavailable
- InterfacesMCP · SDK · STDIO
What it does
AgentAnycast is a peer-to-peer runtime for AI agents that need to be reachable without a public URL. The project positions itself as fixing a gap in agent-to-agent protocols that assume every agent exposes a public endpoint: real agents often run on laptops, behind NAT, or inside corporate networks. It provides NAT traversal (DCUtR hole punching + relay fallback), Noise_XX end-to-end encryption, and skill-based anycast routing so peers discover each other by advertised capabilities (AgentCard + Skill metadata) rather than by DNS or static URLs.
Agent-facing interfaces
- Interaction pattern: SDK + MCP (stdio via agentanycastd or uvx agentanycast-mcp) + auto-managed daemon
- Catalog protocols: MCP, SDK, STDIO
- MCP: ✅ Available — MCP is implemented in the Go daemon (agentanycastd); PyPI package agentanycast-mcp wraps it for uvx workflows.
- Agent Skill: ⚠️ No standalone npx skills add registry entry — onboarding is README + PyPI/npm + MCP docs above.
- Onboarding type:
mcp
How agents onboard
Interaction pattern: SDK + MCP (stdio via agentanycastd or uvx agentanycast-mcp) + auto-managed daemon
pip install agentanycast && agentanycast demo
# Second terminal — use the peer id printed by the demo:
agentanycast send <PEER_ID> "Hello!"
See the full onboarding section in the dossier.
Protocol surface
| Interface | Detail |
|---|---|
| Python SDK | pip install agentanycast — Node, AgentCard, Skill, @node.on_task |
| TypeScript SDK | npm install agentanycast |
| MCP | agentanycastd --mcp-listen stdio or uvx agentanycast-mcp |
| Daemon | ~/.agentanycast/bin/agentanycastd (auto-installed with Python SDK) |