HonorNet agent presentation client (MCP-first)

Make any MCP-capable agent HonorNet-capable. This client holds an agent's credential and its owner's mandate, mints the per-transaction proof of possession, and presents the three-part ADR-009 bundle to the network. The agent decides what the transaction is; this client presents it and relays the network's decision.

What this is — and what it is deliberately not

This is a protocol shim, not an agent framework. The analog is Visa publishing an EMV kernel and a reference terminal: the kernel holds the card's key, mints the cryptogram, and presents it — it decides nothing about what the cardholder buys. (See the four-party model and hard rule 3 in CLAUDE.md.)

The tool surface is closed at exactly three tools:

There is no fourth tool. A tool that makes a purchase decision, ranks options, or calls an LLM is a hard-rule-3 regression and fails CI (tests/test_presentation_client_is_not_an_agent_framework.py).

Design

• Dependency-free. No mcp PyPI package, no LLM SDK, no HTTP library — the MCP stdio transport is a small JSON-RPC 2.0 loop, signing comes from ac-authz (the in-repo reference of the SDK proof/mandate helpers), and transport is stdlib urllib. Same no-supply-chain discipline as ac-decision-record and the AC-620 verifier.
• Issuer-blind. A credential is held and presented unchanged in whatever recognized format it arrives in; no issuer is named in any code path.
• Operator-configured endpoint. Where present POSTs is set by the operator (env), never chosen by the agent in a tool call — the shim never "chooses where to shop."

Run it

# Sign proofs with a known agent key (the credential must be issued to it):
export HONORNET_AGENT_SEED=$(python3 -c "import os;print(os.urandom(32).hex())")

# Where to present (omit for a dry run that returns the bundle only):
export HONORNET_PRESENT_ENDPOINT=http://localhost:8000   # /authorize is appended

python3 -m honornet_presentation.server

The server then speaks MCP over stdin/stdout. Configure it in an MCP host (Claude Desktop, an OpenAI-Agents/LangChain MCP client) as a stdio server:

{
  "mcpServers": {
    "honornet": {
      "command": "python3",
      "args": ["-m", "honornet_presentation.server"],
      "env": {
        "HONORNET_AGENT_SEED": "…64 hex chars…",
        "HONORNET_PRESENT_ENDPOINT": "http://localhost:8000"
      }
    }
  }
}

Library use

from honornet_presentation import PresentationClient

client = PresentationClient(agent_seed=seed, endpoint="http://localhost:8000")
client.load_credential(credential)   # issuer-signed, issued to `seed`'s public key
client.load_mandate(mandate)         # owner-signed
result = client.present(transaction) # {bundle, response: <the network's verdict, relayed>}

Tutorial

The end-to-end walkthrough — off-the-shelf MCP agent → sandbox checkout — is docs/tutorials/agent-mcp-checkout.md.