@tryengrave/mcp

The Model Context Protocol layer for Engrave — attach it to any agentic process and every tool call, input, and output is engraved into a tamper-evident, hash-linked trace. This is the differentiating piece: a black-box recorder for AI agents.

npm install @tryengrave/mcp

Why agents need this

When software takes autonomous action — calling tools, moving money, touching data — "trust me, here are the logs" isn't a defense. Regulators (EU AI Act Art. 12/19, NIST AI RMF, ISO/IEC 42001) increasingly require traceable, integrity-protected records. This package produces exactly that: an append-only, third-party-verifiable trace of what the agent saw, decided, and did.

What's inside

• Tool definitions — tools, callTool: the Engrave MCP tools (record, verify, proof, …) ready to expose to an agent.
• Remote-server core — handleMcpMessage, PROTOCOL_VERSION: a transport-agnostic JSON-RPC handler you can mount behind streamable HTTP.
• Interception proxy (@tryengrave/mcp/proxy) — sit transparently between an agent and any upstream MCP server, engraving every call with byte-fidelity while passing traffic through unchanged.
• Schema — toJsonSchema for tool input/output.

Usage

import { handleMcpMessage } from "@tryengrave/mcp"
import { Engrave } from "@tryengrave/client"

const client = new Engrave({ apiKey: "ek_live_…" })
const response = await handleMcpMessage(client, jsonRpcMessage)

Or wrap an existing MCP server with the CLI — no code changes to the agent:

engrave mcp proxy --upstream stdio:"your-mcp-server" --name your-server

Every intercepted call becomes a provable span in the agent's run trace, which anyone can later verify offline with @tryengrave/verify.

Part of Engrave — the black-box recorder for AI agents. · Agent audit docs · MIT