Signed Audit Trails Recipe

Signed Audit Trails for Claude Code Tool Calls

Cookbook-style walkthrough for cryptographically signed receipts on every Claude Code tool call. This is the teaching skill. For the runtime implementation, install the protect-mcp (opens in a new tab) plugin.

What this gives you

Every tool call (Bash, Edit, Write, WebFetch) is:

1. Evaluated against a Cedar policy before execution. If the policy denies the call, the tool does not run.
2. Signed as an Ed25519 receipt after execution. Receipts are JCS-canonical, hash-chained, and verifiable offline by anyone with the public key.

An auditor, regulator, or counterparty can verify the full chain later with a single CLI command (npx @veritasacta/verify receipts/*.json). No network call, no vendor lookup, no trust in the operator.

When to use the pattern

• Regulated environments (finance, healthcare, critical infrastructure) where you need tamper-evident evidence of agent behavior
• CI/CD pipelines where you want to prove that a policy gate held for every automated build step
• Multi-party collaboration where a counterparty wants to verify your agent’s behavior without trusting your operator
• Compliance contexts (EU AI Act Article 12, SLSA provenance for agent-built software) where standard logging is not sufficient

Step 1: Install the hook configuration

Create .claude/settings.json in your project root:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": ".*",
        "hook": {
          "type": "command",
          "command": "npx protect-mcp@latest evaluate --policy ./protect.cedar --tool \"$TOOL_NAME\" --input \"$TOOL_INPUT\" --fail-on-missing-policy false"
        }
      }
    ],
    "PostToolUse": [
      {
        "matcher": ".*",
        "hook": {
          "type": "command",
          "command": "npx protect-mcp@latest sign --tool \"$TOOL_NAME\" --input \"$TOOL_INPUT\" --output \"$TOOL_OUTPUT\" --receipts ./receipts/ --key ./protect-mcp.key"
        }
      }
    ]
  }
}

The first run of protect-mcp sign generates ./protect-mcp.key (Ed25519 private key) if one does not exist. Commit the public key fingerprint (visible in any receipt’s public_key field); do not commit the private key.

Add the private key and receipt directory to .gitignore:

echo "./protect-mcp.key" >> .gitignore
echo "./receipts/" >> .gitignore

Step 2: Write a Cedar policy

Create ./protect.cedar:

// Allow all read-oriented tools by default.
permit (
    principal,
    action in [Action::"Read", Action::"Glob", Action::"Grep", Action::"WebSearch"],
    resource
);

// Allow Bash commands from a safe list only.
permit (
    principal,
    action == Action::"Bash",
    resource
) when {
    context.command_pattern in [
        "git", "npm", "pnpm", "yarn", "ls", "cat", "pwd",
        "echo", "test", "node", "python", "make"
    ]
};

// Explicit deny on destructive commands. Cedar deny is authoritative.
forbid (
    principal,
    action == Action::"Bash",
    resource
) when {
    context.command_pattern in ["rm -rf", "dd", "mkfs", "shred"]
};

// Restrict writes to the project directory.
permit (
    principal,
    action in [Action::"Write", Action::"Edit"],
    resource
) when {
    context.path_starts_with == "./"
};

Four rules:

• Read-oriented tools always allowed
• Bash allowed for safe command patterns (git, npm, etc.)
• Bash rm -rf and similar destructive commands explicitly denied
• Writes allowed only within the project (./ prefix)

Cedar forbid rules take precedence over permit rules, so destructive commands cannot be bypassed by a later permissive rule.

Step 3: Use Claude Code normally

Start Claude Code. Every tool call goes through both hooks:

You: Please read the README and summarize it.

Claude: I will read README.md.
  [PreToolUse: Read ./README.md -> allow]
  [Tool: Read executes]
  [PostToolUse: receipt rcpt-a8f3c9d2 signed to ./receipts/]

... summary of README ...

A session of 20 tool calls produces 20 receipts, each hash-chained to its predecessor.

Step 4: Inspect a receipt

cat ./receipts/$(ls -t ./receipts/ | head -1)

{
  "receipt_id": "rcpt-a8f3c9d2",
  "receipt_version": "1.0",
  "issuer_id": "claude-code-protect-mcp",
  "event_time": "2026-04-17T12:34:56.123Z",
  "tool_name": "Read",
  "input_hash": "sha256:a3f8c9d2e1b7465f...",
  "decision": "allow",
  "policy_id": "protect.cedar",
  "policy_digest": "sha256:b7e2f4a6c8d0e1f3...",
  "parent_receipt_id": "rcpt-3d1ab7c2",
  "public_key": "4437ca56815c0516...",
  "signature": "4cde814b7889e987..."
}

Every field except signature and public_key is covered by the Ed25519 signature. Modifying any field after signing invalidates the signature.

Step 5: Verify the receipt chain

npx @veritasacta/verify ./receipts/*.json

Exit codes:

Step 6: Demonstrate tamper detection

Modify any receipt’s decision field from allow to deny:

python3 -c "
import json, os
path = './receipts/' + sorted(os.listdir('./receipts'))[-1]
r = json.loads(open(path).read())
r['decision'] = 'deny'
open(path, 'w').write(json.dumps(r))
"

npx @veritasacta/verify ./receipts/*.json

The verifier exits with code 1 and reports which receipt failed. The Ed25519 signature no longer matches the JCS-canonical bytes of the tampered payload.

Restore the field and verification passes again.

How the cryptography works

Three invariants make receipts verifiable offline across any conformant implementation:

1. JCS canonicalization (RFC 8785) before signing. Keys sorted, whitespace minimized, strings NFC-normalized. Two independent implementations produce byte-identical signing payloads for the same receipt content.
2. Ed25519 signatures (RFC 8032) over the canonical bytes. Deterministic, fixed-size, no nonce dependency.
3. Hash chain linkage. Each receipt’s parent_receipt_hash is the SHA-256 of the predecessor’s canonical form. Insertions, deletions, and reorderings break later receipts.

For the formal wire format see draft-farley-acta-signed-receipts (opens in a new tab).

Cross-implementation interop

The receipt format has four independent implementations today:

A receipt produced by any of them verifies against @veritasacta/verify (opens in a new tab). The auditor does not need to trust the operator’s tooling choice: the format is the contract.

CI/CD integration

Gate merges on receipt chain verification so no build lands with a broken evidence chain:

# .github/workflows/verify-receipts.yml
name: Verify Decision Receipts
on: [push, pull_request]

jobs:
  verify:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with: { node-version: '20' }
      - name: Run governed agent
        run: python scripts/run_agent.py > receipts.jsonl
      - name: Verify receipt chain
        run: npx @veritasacta/verify receipts.jsonl

Archive the receipts as an artifact so the chain survives beyond the job run:

      - name: Upload receipts
        if: always()
        uses: actions/upload-artifact@v4
        with:
          name: decision-receipts
          path: receipts/

Composition with SLSA provenance for agent-built software

When Claude Code builds and releases software (running npm install, npm build, npm publish as tool calls), the receipt chain is the per-step build log. SLSA Provenance v1 has an extension point for this: the byproducts field can reference the receipt chain alongside the build attestation.

The agent-commit build type (opens in a new tab) documents the pattern using the ResourceDescriptor shape:

{
  "name": "decision-receipts",
  "digest": { "sha256": "..." },
  "uri": "oci://registry/org/build-xyz/receipts:sha256-...",
  "annotations": {
    "predicateType": "https://veritasacta.com/attestation/decision-receipt/v0.1",
    "signerRole": "supervisor-hook"
  }
}

The SLSA provenance is signed by the builder identity; the receipt attestation is signed by the supervisor-hook identity. Two trust domains, cross-referenced at the byproduct layer. See slsa-framework/slsa#1594 (opens in a new tab) for the composition discussion.

Common pitfalls

Private key in version control. The generated ./protect-mcp.key must not be committed. The examples above add it to .gitignore. If a key is accidentally committed, rotate immediately (delete the key file and let the hook regenerate on next run).

Hook command quoting. The hooks receive $TOOL_NAME and $TOOL_INPUT as environment variables. Keep the quoting "$TOOL_INPUT" so inputs with spaces or special characters pass through intact.

Receipts directory in CI. If Claude Code runs in CI, upload receipts as an artifact at the end of the job or the chain is lost at job end.

Policy is missing. The example PreToolUse hook uses --fail-on-missing-policy false so an absent ./protect.cedar does not break Claude Code out of the box. Remove this flag in production so a missing policy is treated as a hard failure.

Related in this marketplace

• protect-mcp (opens in a new tab) — the runtime hook implementation (use this plugin in production)
• review-agent-governance (opens in a new tab) — require human approval before review-surface actions; composes with protect-mcp

References

• draft-farley-acta-signed-receipts (opens in a new tab) — IETF draft, receipt wire format
• RFC 8032 (opens in a new tab) — Ed25519
• RFC 8785 (opens in a new tab) — JCS
• Cedar policy language (opens in a new tab)
• protect-mcp on npm (opens in a new tab)
• @veritasacta/verify on npm (opens in a new tab)
• in-toto/attestation#549 (opens in a new tab) — Decision Receipt predicate proposal
• agent-commit build type (opens in a new tab) — SLSA provenance for agent-produced commits
• Microsoft Agent Governance Toolkit (opens in a new tab) (examples/protect-mcp-governed/)
• AWS Cedar for Agents (opens in a new tab)