Introduction

The lockfile for the Agentic Web.


The lockfile for the Agentic Web.

MCPTrust secures AI agents by cryptographically verifying MCP servers before use.

The Problem: AI Agents Have a Blind Trust Problem

AI agents (Claude, GPT, Gemini) connect to external tools via the Model Context Protocol (MCP). These tools can do anything: read files, execute code, send emails.

The Risk: Tools can change silently. A harmless read_file tool today could become exec_shell tomorrow.

Real-world analogy: "Would you install an npm package without a lockfile? Then why let your AI agent use tools without one?"

The Solution: Scan → Lock → Enforce

  1. Scan: Interrogate the MCP server to discover all its tools, their schemas, and risk levels.
  2. Lock: Create an immutable mcp-lock.json that captures the exact shape of the server.
  3. Enforce: Before every use, verify the server hasn't drifted from the locked state.
# 1. Scan a server
mcptrust scan -- "npx -y @modelcontextprotocol/server-filesystem /tmp"
 
# 2. Lock it
mcptrust lock -- "npx -y @modelcontextprotocol/server-filesystem /tmp"
 
# 3. Enforce at runtime
mcptrust proxy --lock mcp-lock.json -- "npx -y @modelcontextprotocol/server-filesystem /tmp"

Key Features

FeatureDescription
🔒 Drift DetectionDetects any change in tool description, schema, or risk level
📜 Policy EnforcementDefine rules in CEL (Common Expression Language)
✍️ Cryptographic SigningEd25519 or Sigstore keyless signatures
🔗 Supply Chain VerificationSLSA provenance and npm integrity checks
🛡️ Runtime ProxyStdio middleware that blocks unapproved tools

Who This Is For

  • Security teams deploying AI agents in production
  • Developers building AI-powered applications
  • Enterprises with compliance/SOC2/FedRAMP requirements
  • Anyone using Claude Desktop, LangChain, or custom MCP integrations