protect-mcp-governance

name	protect-mcp-governance
description	Agent governance skill for MCP tool calls — Cedar policy authoring, shadow-to-enforce rollout, and Ed25519 receipt verification.
risk	safe
source	community
source_repo	scopeblind/scopeblind-gateway
source_type	official
date_added	2026-04-05

MCP Agent Governance with protect-mcp

Overview

Guidance for governing AI agent tool calls using Cedar policies and Ed25519 signed receipts. This skill teaches how to write access-control policies for MCP servers, run them in shadow mode for observation, and verify the cryptographic audit trail.

When to Use This Skill

• Use when you need to control which MCP tools an agent can call and under what conditions
• Use when you want a tamper-evident audit trail for agent tool executions
• Use when rolling out governance policies gradually (shadow mode first, then enforce)
• Use when authoring Cedar policies for MCP tool access control
• Use when verifying that a receipt or audit bundle has not been tampered with

Do Not Use This Skill

• When you need general application security auditing (use @security-auditor)
• When you need to scan code for vulnerabilities (use @security-audit)
• When you need compliance framework guidance without agent-specific governance

How It Works

protect-mcp intercepts MCP tool calls, evaluates them against Cedar policies (the same policy engine used by AWS Verified Permissions), and signs every decision as an Ed25519 receipt. The receipt is a cryptographic proof that a specific policy was evaluated against a specific tool call at a specific time.

July 2026 protocol notes (MCP 2026-07-28)

FidusGate’s gateway is dual-era:

• stdio — legacy initialize handshake retained for Cursor/local clients (2025-11-25 preferred; 2024-11-05 accepted).
• HTTP POST /mcp — stateless Streamable HTTP: require MCP-Protocol-Version, Mcp-Method, and (for named methods) Mcp-Name. Header/body disagreement is rejected before Cedar runs.
• Modern clients should call server/discover instead of relying on session init.
• Demo OAuth surface: GET /.well-known/oauth-protected-resource (RFC 9728); map residual risks via OWASP MCP Top 10 ADR.
• Operator guide: mcp-2026-07-28-migration.md.

Agent → protect-mcp → Cedar policy evaluation → MCP Server
                ↓
        Ed25519 signed receipt

Three modes of operation:

1. Shadow mode (default) — logs decisions without blocking. Use this to observe what your policies would do before enforcing them.
2. Enforce mode — blocks tool calls that violate policy. Use after shadow-mode validation.
3. Hooks mode — integrates with Claude Code hooks for pre/post tool-call governance.

Core Concepts

Cedar Policies

Cedar is a policy language designed for authorization. Policies are evaluated locally via WASM — no network calls required.

// Allow read-only file operations
permit(
  principal,
  action == Action::"call_tool",
  resource
) when {
  resource.tool_name in ["read_file", "list_directory", "search_files"]
};

// Deny destructive operations
forbid(
  principal,
  action == Action::"call_tool",
  resource
) when {
  resource.tool_name in ["execute_command", "delete_file", "write_file"]
  && resource has args
  && resource.args.contains("rm -rf")
};

Risk-Tiered Development Governance

To mitigate the four main latent development risks (Runaway Concurrency, Self-Authoring Rot, Yak-Shaving loops, and Micro-Auditing lockups), we map agentic tools to four explicit risk tiers:

Tier	Name	Development Tools Affected	Cedar Governance Mode
Tier 1	Low Risk	read_file, list_directory, grep_search	Auto-Approve (No restrictions)
Tier 2	Medium Risk	write_file, replace_file_content outside src/	Shadow-to-Enforce (Alerts on directory drift)
Tier 3	High Risk	execute_command (Parallel Worker spawns, skill-creator)	Interactive Authorization (Requires --yes flags)
Tier 4	Critical Risk	Unsandboxed network downloads, global policy deletes	Strict Interdiction (Pre-authorized receipt required)

Concrete Cedar Policies for Development De-Risking

// 1. TIER 3: Prevent runaway token burn from parallel workers unless explicitly approved
forbid(
  principal,
  action == Action::"call_tool",
  resource
) when {
  resource.tool_name == "execute_command"
  && resource.args.contains("orchestrate-batch-refactor")
  && !resource.args.contains("--yes")
};

// 2. TIER 3: Block dynamic, unverified skill creation to prevent Prompt Pollution
forbid(
  principal,
  action == Action::"call_tool",
  resource
) when {
  resource.tool_name == "execute_command"
  && resource.args.contains("skill-creator")
  && !resource.args.contains("--manual-auth")
};

// 3. TIER 2: Enforce a timeout block on loop-prone environment/Docker setups
permit(
  principal,
  action == Action::"call_tool",
  resource
) when {
  resource.tool_name in ["execute_command", "write_file"]
  && resource.args.contains("devcontainer-setup")
} when {
  // Shadow-observe execution state to abort loops after 3 failures
  context.session.failure_count < 3
};

Signed Receipts

Every policy decision produces a signed receipt:

{
  "payload": {
    "type": "protectmcp:decision",
    "tool_name": "read_file",
    "decision": "allow",
    "policy_digest": "sha256:9d0fd4c9e72c1d5d",
    "issued_at": "2026-04-05T14:32:04.102Z",
    "issuer_id": "sb:issuer:de073ae64e43"
  },
  "signature": {
    "alg": "EdDSA",
    "kid": "sb:issuer:de073ae64e43",
    "sig": "2a3b5022..."
  }
}

The receipt format follows IETF Internet-Draft draft-farley-acta-signed-receipts.

Step-by-Step Guide

1. Initialize Governance for a Project

# Install and initialize hooks (Claude Code integration)
npx protect-mcp init-hooks

# Or run as a standalone MCP gateway
npx protect-mcp serve

This creates a protect-mcp.config.json and a starter Cedar policy in your project root.

2. Write Your First Policy

Create policy.cedar in your project:

// Start permissive — allow everything in shadow mode
permit(
  principal,
  action == Action::"call_tool",
  resource
);

3. Run in Shadow Mode (Observe First)

# Shadow mode is the default — logs decisions without blocking
npx protect-mcp --policy policy.cedar -- node your-mcp-server.js

Review the shadow log to understand what your agent is doing before writing restrictive policies.

4. Tighten and Enforce

Once you understand the tool-call patterns, write specific policies:

// Allow file reads, deny writes outside src/
permit(
  principal,
  action == Action::"call_tool",
  resource
) when {
  resource.tool_name == "read_file"
};

permit(
  principal,
  action == Action::"call_tool",
  resource
) when {
  resource.tool_name == "write_file"
  && resource has args
  && resource.args.path like "src/*"
};

// Deny everything else
forbid(
  principal,
  action == Action::"call_tool",
  resource
);

Switch to enforce mode:

npx protect-mcp --policy policy.cedar --enforce -- node your-mcp-server.js

5. Verify Receipts

# Verify a single receipt
npx @veritasacta/verify receipt.json --key <public-key-hex>

# Verify an audit bundle (multiple receipts + keys)
npx @veritasacta/verify bundle.json --bundle

# Self-test the verifier (proves it works offline)
npx @veritasacta/verify --self-test

Exit codes: 0 = signature valid (proven authentic), 1 = signature invalid (proven tampered), 2 = verifier error (malformed input).

Examples

Example 1: Governance for a Claude Code Session

# Initialize hooks
npx protect-mcp init-hooks

# Claude Code now generates a signed receipt for every tool call.
# Receipts are stored in .protect-mcp/receipts/

Explanation: After initialization, every tool call Claude Code makes is logged with a signed receipt. No tool calls are blocked (shadow mode).

Example 2: Restrict a Production MCP Server

// Only allow approved tools with rate limiting
permit(
  principal,
  action == Action::"call_tool",
  resource
) when {
  resource.tool_name in [
    "get_customer",
    "search_orders",
    "list_products"
  ]
};

forbid(
  principal,
  action == Action::"call_tool",
  resource
) when {
  resource.tool_name in [
    "delete_customer",
    "modify_payment",
    "execute_sql"
  ]
};

Explanation: A production MCP server that serves customer data. Read-only operations are permitted; destructive operations are blocked.

Example 3: Verify an Audit Bundle After an Incident

# Export the session's audit bundle
npx protect-mcp export-bundle --session sess_abc123 --out audit.json

# Verify every receipt in the bundle
npx @veritasacta/verify audit.json --bundle

# Expected output:
# ✓ Bundle: VALID
#   Total:    47
#   Passed:   47
#   Failed:   0

Explanation: After an incident, export the audit bundle and verify that no receipts have been tampered with. The bundle contains all receipts from the session plus the signing keys needed for verification.

Best Practices

• ✅ Do: Start in shadow mode and observe before enforcing
• ✅ Do: Use policy_digest to track which policy version produced each decision
• ✅ Do: Store receipts alongside your application logs for correlation
• ✅ Do: Pin the verifier version when integrating into CI (@veritasacta/verify@0.2.5)
• ❌ Don't: Skip shadow mode and go straight to enforce in production
• ❌ Don't: Trust claimed_issuer_tier without independent verification
• ❌ Don't: Treat a valid signature as proof the signer is trustworthy — it only proves the receipt has not been tampered with since signing

Troubleshooting

Problem: Receipts fail verification with no_public_key

Symptoms: npx @veritasacta/verify receipt.json returns exit 2 with no_public_key Solution: Provide the public key explicitly: --key <64 hex chars>. The receipt does not embed the public key by default. Check protect-mcp.config.json for the issuer's public key.

Problem: Shadow mode shows unexpected denials

Symptoms: Shadow log shows deny decisions for tools you expected to be allowed Solution: Check your Cedar policy ordering. Cedar evaluates forbid rules before permit rules — a broad forbid will override specific permit rules.

Problem: Enforce mode blocks a legitimate tool call

Symptoms: Agent reports a tool call was denied after switching to enforce mode Solution: Add the tool to your permit policy or switch back to shadow mode: remove --enforce flag. Review the receipt's deny_reason field for the specific policy violation.

Related Skills

• @security-auditor — General security auditing and compliance
• @security-audit — Code vulnerability scanning
• @mcp-development — MCP server development patterns

Additional Resources

• protect-mcp on npm — MIT licensed
• Cedar Policy Language — AWS open-source policy engine
• IETF Draft: Signed Receipts — Receipt format specification
• @veritasacta/verify — Apache-2.0 verifier, works offline

Limitations

• Use this skill only when the task clearly matches the scope described above.
• Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
• Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.