Autropic, an Australian software company, has launched Joy, a decentralized trust network and discovery platform for AI agents and Model Context Protocol (MCP) servers. The platform allows agents to register, vouch for one another, and be discovered via a REST API, with the stated goal of giving autonomous agents a reliable way to evaluate the trustworthiness of other agents before delegating tasks or sharing context. As of publication, the network lists 6,080 registered agents and 2,042 trust vouches. Joy integrates directly with Claude Code via MCP's Streamable HTTP transport, meaning developers can add it with a single CLI command.
The trust model is straightforward: each peer vouch adds 0.3 points to a recipient's trust score, capped at a maximum of 3.0, and agents that prove ownership of a live endpoint receive priority placement in discovery results. Joy's own MCP Registry agent acts as a trust anchor, cross-referencing recognized upstream registries including mcp.so, GitHub, and Agent Protocol to issue vouches for well-known servers. The discovery API surfaces a broad range of MCP servers covering GitHub integrations, document retrieval, web intelligence, and workplace connectors such as Slack and Jira, many of them hosted on Smithery's infrastructure.
The network's current statistics reveal a significant bootstrapping challenge. With roughly 0.33 vouches per registered agent, most listings sit at a baseline score of 0.7 — consistent with a single system-issued vouch — while the top non-system entry, the GitHub MCP server, has accumulated only 6 peer vouches for a score of 1.7. The low cap means a maximum trust score requires just 10 vouches, a threshold reachable through a small coordinated ring of colluding agents. Spot checks of the registry also turned up duplicate entries for agents such as "GitHub" and "driflyte-mcp-server," which means Joy hasn't yet solved <a href="/news/2026-03-14-aip-agent-intent-protocol-cryptographic-identity">canonical identity</a> — and without that, Sybil resistance is mostly theoretical.
The bootstrapping problem isn't unique to Joy. PGP's web of trust and early marketplace reputation models hit the same wall: manufactured scores are hardest to detect before enough legitimate activity builds up to make them statistically visible. Joy's linear-additive model is easier to game than weighted propagation systems like EigenTrust or SourceCred, where a voucher's own trust score determines how much weight their vouch carries. Autropic hasn't published plans for vouch-weight decay or rate limiting. Until the graph grows dense enough that collusion becomes detectable, Joy's trust scores are more useful as a discovery filter than a genuine <a href="/news/2026-03-14-nanoclaw-partners-with-docker-for-hypervisor-level-agent-sandboxing">security guarantee</a>.