Kanoniv

The identity and delegation layer for AI agents.

---

Kanoniv gives every AI agent a cryptographic identity (did:agent:) and verifiable delegation chains. Agents carry proof of who they are and what they are authorized to do. Any MCP server verifies the proof in 5 lines. No external lookups needed.

Architecture

Layer 0: Identity           did:agent: DIDs              (open source)
Layer 1: Delegation         attenuated authority chains   (open source)
Layer 2: MCP Auth           proofs on tool calls          (open source)
Layer 3: Memory             entity-linked knowledge       (Cloud)
Layer 4: Agent Resolution   same agent across platforms   (Cloud)
Layer 5: Audit              signed provenance trail       (Cloud)

Layers 0-2 are free and open source (MIT). Layers 3-5 require Kanoniv Cloud.

Repositories

Repo	Description
agent-auth	Core library: Ed25519 identity, delegation, MCP proofs. Rust + TypeScript + Python.
kanoniv-crewai	CrewAI integration: DelegatedCrew, delegated_tool
kanoniv-langgraph	LangGraph integration: DelegatedGraph, @delegated_node
kanoniv-autogen	AutoGen integration: AuthorityManager, DelegatedAgent
kanoniv-openai-agents	OpenAI Agents SDK integration: DelegatedRunner, @delegated_tool

Install

# Core library
cargo add kanoniv-agent-auth        # Rust
npm install @kanoniv/agent-auth     # TypeScript
pip install kanoniv-agent-auth      # Python

# CLI
npx @kanoniv/agent-auth generate
npx @kanoniv/agent-auth delegate --to did:agent:... --scope resolve,search --max-cost 5
npx @kanoniv/agent-auth verify --proof proof.json --root did:agent:...

# Framework integrations
pip install kanoniv-crewai
pip install kanoniv-langgraph
pip install kanoniv-autogen
pip install kanoniv-openai-agents

Quick Start (5 lines to add auth to any MCP server)

import { McpProof, verifyMcpCall } from "@kanoniv/agent-auth";

function handleToolCall(args) {
  const { proof, cleanArgs } = McpProof.extract(args);
  if (proof) {
    const result = verifyMcpCall(proof, rootIdentity);
    console.log(`Agent ${result.invoker_did} verified (depth: ${result.depth})`);
  }
}

How Delegation Works

Root Authority (Human)
  |-- delegates to Manager: [search, resolve, merge], max $50, expires 24h
      |-- delegates to Worker: [search], max $10
          |-- calls MCP tool with proof
              |-- server verifies chain, checks caveats, executes or rejects

Authority narrows at each step. Caveats accumulate. A sub-agent can never exceed its parent's scope.

Documentation

• Getting Started
• Identity (Layer 0)
• Delegation (Layer 1)
• MCP Auth (Layer 2)
• Memory (Layer 3)
• Agent Resolution (Layer 4)
• Security Model
• SDK References

MCP Spec Proposal

We proposed adding an auth field to MCP tool calls for agent delegation proofs: Discussion #2404

License

MIT

---

Website · Docs · Core Library · Pricing