What it looks like

After a session where the agent reads files, runs a command, browses a page, and writes output, querying the audit trail returns:

{
  "total_receipts": 5,
  "total_chains": 1,
  "by_risk": { "low": 4, "high": 1 },
  "by_status": { "success": 4, "failure": 1 },
  "by_action": {
    "filesystem.file.read": 2,
    "filesystem.file.create": 1,
    "system.command.execute": 1,
    "system.browser.navigate": 1
  },
  "results": [
    { "id": "rec-…01", "timestamp": "2026-04-01T02:10:01Z", "action": "filesystem.file.read",    "risk": "low",  "target": "read_file",        "status": "success", "sequence": 1 },
    { "id": "rec-…02", "timestamp": "2026-04-01T02:10:02Z", "action": "filesystem.file.read",    "risk": "low",  "target": "read_file",        "status": "failure", "sequence": 2 },
    { "id": "rec-…03", "timestamp": "2026-04-01T02:10:03Z", "action": "system.command.execute",  "risk": "high", "target": "run_command",      "status": "success", "sequence": 3 },
    { "id": "rec-…04", "timestamp": "2026-04-01T02:10:04Z", "action": "system.browser.navigate", "risk": "low",  "target": "browser_navigate", "status": "success", "sequence": 4 },
    { "id": "rec-…05", "timestamp": "2026-04-01T02:10:05Z", "action": "filesystem.file.create",  "risk": "low",  "target": "write_file",       "status": "success", "sequence": 5 }
  ]
}

Verifying the chain confirms nothing was tampered with:

Chain "chain_openclaw_main_sid-42" is valid: 5 receipts, all signatures and hash links verified.

Every receipt is a signed W3C Verifiable Credential — parameters are hashed by default (with optional plaintext disclosure via parameterDisclosure), and each receipt is hash-linked to the previous one, forming a tamper-evident chain.

Why receipts?

AI agents that read files, run commands, and browse the web are powerful — but that power needs accountability. When an agent operates autonomously, you need to know exactly what it did, prove that the record hasn't been tampered with, and keep sensitive details private.

Use cases:

• Post-incident review — your agent ran overnight and something broke. The receipt chain shows exactly which commands it ran, in what order, and whether each succeeded or failed — with cryptographic proof that the log hasn't been altered after the fact.
• Compliance and audit — regulated environments require evidence of what systems did and why. Receipts are W3C Verifiable Credentials with Ed25519 signatures, giving auditors a tamper-evident trail they can independently verify.
• Safer autonomous agents — the agent can query its own audit trail mid-session. Before taking a high-risk action, it can check what it has already done and whether previous steps succeeded, enabling self-correcting workflows.
• Multi-agent trust — when agents collaborate, receipts serve as proof of prior actions. Agent B can verify that Agent A actually completed step 1 before proceeding to step 2, without trusting a shared log.
• Cost and usage tracking — every tool call is classified by type and risk level, giving you a structured breakdown of what your agent spent its time on across sessions.

Beyond local storage

Today, receipts are stored locally in SQLite — fully under your control. The Agent Receipts protocol is designed for receipts to travel further when you choose: publishing to a shared ledger, forwarding to a compliance system, or exchanging with other agents as proof of prior actions. The receipts are portable W3C Verifiable Credentials, but where they go is always your decision.

How it works

Every time the OpenClaw agent executes a tool, this plugin:

1. Classifies the action using the Agent Receipts taxonomy
2. Creates a signed receipt — a W3C Verifiable Credential with Ed25519 proof
3. Hash-links it into a per-session chain (tamper-evident)
4. Stores it in a local SQLite database

The agent also gets two introspection tools to query and verify its own audit trail.

OpenClaw Gateway
  │
  ├─ before_tool_call ──► capture params + timing
  │
  ├─ [tool executes]
  │
  └─ after_tool_call ──► classify → sign → chain → store

Install

openclaw plugins install @agnt-rcpt/openclaw

Then enable the plugin in your OpenClaw config. See docs/INSTALL.md for tool visibility setup and configuration options.

CLI — Receipt Explorer

Query and verify receipts outside of agent sessions, useful for auditing and debugging.

# List all receipts
npx @agnt-rcpt/openclaw receipts

# Filter by risk level
npx @agnt-rcpt/openclaw receipts --risk high

# Filter by action type and output as JSON
npx @agnt-rcpt/openclaw receipts --action system.command.execute --json

# `receipts` always returns a collection — use `export --id` to fetch a single receipt by ID
npx @agnt-rcpt/openclaw export --id urn🧾abc-123

# Filter receipts --json output with jq (fields: id, action, risk, target, status, sequence, chain_id, timestamp)
npx @agnt-rcpt/openclaw receipts --json \
  | jq '.receipts[] | select(.risk == "high" and .action == "system.command.execute")'

# Verify all chains
npx @agnt-rcpt/openclaw verify

# Verify a specific chain
npx @agnt-rcpt/openclaw verify --chain chain_openclaw_main_sid-42

# Export a chain as JSON-LD (full W3C Verifiable Credentials)
npx @agnt-rcpt/openclaw export --chain chain_openclaw_main_sid-42

# Export as a W3C Verifiable Presentation envelope
npx @agnt-rcpt/openclaw export --chain chain_openclaw_main_sid-42 --format presentation

Note: parameterDisclosure controls what gets stored inside receipts — it does not add fields to receipts --json output. To inspect parameters_disclosure values, export the full receipt with export --id or export --chain. See Parameter disclosure for configuration details.

Run npx @agnt-rcpt/openclaw --help for all options including --status, --limit, and --db.

Agent tools

ar_query_receipts

Search the audit trail by action type, risk level, or outcome status. Returns receipt summaries and aggregate statistics.

> Query all high-risk actions from this session

{
  "total_receipts": 12,
  "results": [
    { "action": "filesystem.file.delete", "risk": "high", "target": "delete_file", "status": "success", "sequence": 7 },
    { "action": "system.command.execute", "risk": "high", "target": "run_command", "status": "success", "sequence": 3 }
  ]
}

ar_verify_chain

Cryptographically verify the integrity of the receipt chain. Checks Ed25519 signatures, hash links, and sequence numbering.

> Verify the audit trail for this session

Chain "chain_openclaw_main_sid-42" is valid: 12 receipts, all signatures and hash links verified.

What's in a receipt?

Each receipt is a W3C Verifiable Credential signed with Ed25519, recording:

Taxonomy

The plugin maps OpenClaw tool names to Agent Receipts action types:

See taxonomy.json for the full 20-tool mapping. Override with a custom file via the taxonomyPath config option.

Configuration

All settings are optional — the plugin works out of the box with sensible defaults.

Default config block:

{
  "plugins": {
    "entries": {
      "openclaw-agent-receipts": {
        "config": {
          "enabled": true,
          "dbPath": "~/.openclaw/agent-receipts/receipts.db",
          "keyPath": "~/.openclaw/agent-receipts/keys.json",
          // "taxonomyPath": "/path/to/custom-taxonomy.json",  // optional — overrides bundled taxonomy
          "parameterDisclosure": false  // false | true | "high" | string[]
        }
      }
    }
  }
}

Ed25519 signing keys are generated automatically on first run and persisted to keyPath.

Parameter disclosure

By default, action parameters are hashed but not stored in plaintext. Enable parameterDisclosure to selectively disclose specific fields per action type — useful for auditing high-risk commands without exposing sensitive data on lower-risk calls.

{
  "plugins": {
    "entries": {
      "openclaw-agent-receipts": {
        "config": {
          "parameterDisclosure": "high"
        }
      }
    }
  }
}

Options:

With "high" enabled, a system.command.execute receipt includes:

{
  // ...other receipt fields
  "parameters_hash": "sha256:9c84a8c9...",
  "parameters_disclosure": {
    "command": "echo \"Testing agent-receipts plugin fix\""
  }
}

The hash always covers the full original parameters regardless of disclosure config. Only the first matching field from the taxonomy's disclosure_fields list is included in parameters_disclosure, and non-string values are JSON-stringified. Disclosed values are signed and durable — do not list fields that may contain secrets.

Project structure

src/
  index.ts          # Plugin entry — wires hooks, tools, service
  cli.ts            # Receipt Explorer CLI (npx @agnt-rcpt/openclaw)
  hooks.ts          # before_tool_call / after_tool_call → receipt creation
  classify.ts       # Tool name → action type + risk level classification
  chain.ts          # Per-session hash-linked chain state
  tools.ts          # ar_query_receipts + ar_verify_chain
  config.ts         # Config resolution + Ed25519 key management
taxonomy.json       # Default OpenClaw tool → action type mappings

Development

pnpm install
pnpm test              # run the test suite
pnpm run typecheck     # TypeScript strict mode
pnpm test:coverage     # with V8 coverage

Ecosystem

License

MIT