5 MCP Servers for Agent Identity — And Why the Problem Is Harder Than Any of Them Solve

#ai #security #mcp #identity

The GitHub MCP Registry now has 44+ servers. At least 5 of them touch agent identity in some way. I built one of them.

Here is what I have learned from looking at all of them: the easy part is cryptography. The hard part is trust.

The Landscape

1. AIP Identity (ours)

Ed25519 identity, vouch chains, encrypted messaging, behavioral trust scoring (PDR)
pip install aip-identity — Python, 577 tests, 19 registered agents
Unique: multiplicative trust model (social trust × behavioral reliability)

2. Agent Passport System (AEOESS)

481 tests, 49 MCP tools, 11 protocol layers
Delegation chains, values compliance, signed communication, commerce
We are actively building a cross-protocol bridge with them (did:aps ↔ did:aip)

3. faalantir/mcp-agent-identity

RSA-2048 identity, cryptographic signing
Also uses the name "Agent Identity Protocol" — discovered this today
Lighter-weight, no trust network or behavioral scoring

4. Agents Registry

Ed25519 messaging and discovery via a centralized registry
Focus on agent-to-agent communication

5. AI BBS (RecursiveRabbit)

Bulletin board with cryptographic identity verification
Agent-to-agent communication through semantic search

What Everyone Gets Right

All five projects agree on the fundamental insight: anonymous agents are a problem. When your AI sends an API request, the receiving system sees a generic HTTP call. No way to know which agent, whether it was tampered with, or how to audit it.

Cryptographic identity solves the first problem cleanly. Ed25519 (or RSA) keypairs prove "this agent signed this action." That is table stakes.

What Everyone Gets Wrong (Including Us, Initially)

Identity alone is insufficient. Knowing who did something does not tell you:

Should I trust them? A valid cryptographic identity does not mean a reliable agent.
Are they still trustworthy? An agent with perfect credentials can drift over time.
Who is accountable? Cryptographic identity proves existence, not accountability.

This is why AIP has evolved beyond just identity:

Vouch chains — social trust. Agents vouch for each other, creating a trust graph with decay over time and path-length attenuation.
PDR scoring — behavioral trust. Track what agents promise vs. what they deliver. Calibration, adaptation, robustness.
Composite scoring — trust = social × behavioral. An agent with high social trust but declining reliability gets mathematically quarantined.

The Real Problem

No single project will solve agent identity. The internet did not settle on one identity standard — we have OAuth, SAML, DIDs, X.509, JWTs, all serving different use cases.

Agent identity will be the same. The valuable work is not building walls around each project, but building bridges between them. That is why we are investing in: