Yabbie Net

A safety net for AI agents. Catches unsafe tool calls before they execute.

Yabbie Net — A Safety Net for AI Agents

Yabbie Net is an open-source MCP proxy that sits between your OpenClaw agent and its tools, catching unsafe actions before they execute.

Source code: github.com/Devlines/yabbie.net (MIT licensed, fully auditable)

What It Does

Three tiers of protection:

1. Deterministic rules (instant, free, no external calls) — file path deny lists, tool blocklists, rate limits
2. AI intent judge (optional, opt-in only) — a small model checks if each action matches your stated goal
3. Human escalation (rare) — you only see actions that are uncertain AND irreversible

Tier 1 requires no API keys, makes no network calls, and sends no data externally. It runs entirely locally using pattern matching.

Security and Privacy

Because this is a security tool, transparency matters. Here is exactly what Yabbie Net does and does not do:

What the proxy sees:

• Tool names and arguments passing through the MCP stdio channel (same data the MCP server already receives)

What is sent externally (ONLY when tier2 is explicitly enabled):

• Tool name
• Truncated argument summary (keys + types + short values; large content like file bodies is replaced with byte counts — never sent in full)
• Your taskContext string from yabbie.yaml
• Sent to: Anthropic API (if provider: anthropic) or localhost Ollama (if provider: ollama)

What is NEVER sent externally:

• Full file contents, full argument values, or raw MCP traffic
• Any data when tier2 is disabled (the default)
• Telemetry is opt-in only (telemetry: true in config). Anonymous aggregates only (latency percentiles, verdict ratios). No tool names, no arguments, no file paths.

Credentials:

• ANTHROPIC_API_KEY — required ONLY if you enable tier2 with provider: anthropic. Set it in your shell environment. Not needed for tier1-only mode.
• No other credentials are required.

Setup

Install as a project dependency (recommended over global install):

bash

npm install yabbie-net@0.2.0

Or install globally:

bash

npm install -g yabbie-net@0.2.0

In your openclaw.json, wrap any MCP server:

json

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["yabbie-net@0.2.0", "--verbose", "--", "npx", "-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
    }
  }
}

Note: This modifies how your MCP servers are invoked. Yabbie acts as a transparent proxy — all tool calls pass through it. Review the source code to understand the interception mechanism.

Configuration

Create yabbie.yaml in your project root:

yaml

version: 1

# Tier 1: Local rules only. No network calls. No API keys needed.
tier1:
  files:
    deny: ["**/.env", "**/secrets*", "**/*.pem", "**/*.key"]
  tools:
    deny: ["shell_exec"]
  rateLimit:
    maxCallsPerMinute: 30

# Tier 2: DISABLED by default. Requires ANTHROPIC_API_KEY if enabled.
# Sends tool name + truncated args to the configured provider.
tier2:
  enabled: false
  provider: anthropic       # "anthropic" (requires ANTHROPIC_API_KEY) or "ollama" (local, no key)
  sensitivity: balanced
  taskContext: "Describe what your agent should be doing"

# Tier 3: Human approval for uncertain + irreversible actions.
tier3:
  enabled: true
  channel: stderr

log:
  verbose: true

# Anonymous telemetry. Off by default. No tool names or arguments ever sent.
telemetry: false

Recommended first step: Start with tier1 only (the default). No API keys needed, no external calls. Add tier2 later once you've reviewed the audit logs and understand your agent's patterns.

How It Affects Other Skills

When you route an MCP server through Yabbie in openclaw.json, the proxy intercepts tools/call JSON-RPC messages for that server. This means:

• Tool calls may be blocked if they match deny rules (the agent receives a clear error message)
• Tool calls may be delayed by ~1ms for tier1 checks, or ~300-800ms if tier2 is enabled
• All other MCP messages (initialize, tools/list, notifications) pass through unmodified
• The proxy does not modify tool arguments or responses — it only allows or blocks

Why This Exists

Cisco research found 36% of ClawHub skills contain prompt injection vulnerabilities. Agents have exfiltrated data, created unauthorized accounts, and deleted production databases. NVIDIA's NemoClaw addresses this but requires RTX/DGX hardware.

Yabbie Net is a lightweight, software-only safety layer that runs anywhere.

Audit and Verify

All actions are logged locally to .yabbie/audit.jsonl:

bash

npx yabbie-net log --tail 20   # View recent actions
npx yabbie-net stats            # View performance metrics

Links

• Source Code — fully auditable, MIT licensed
• Report Issues