agent-looker-for-claude-cowork

Agent Looker - Claude Cowork Plugin

A plugin for Claude Cowork that protects AI agents from unsafe URLs, malicious content, and prompt injection attacks via the Agent Looker MCP server.

What it does

Agent Looker adds two layers of protection to your Claude Cowork sessions:

Hooks (automatic) -- System-level guards that inject security rules before and after web tool calls:

• PreToolUse -- Before every tool call, security rules are injected into Claude's context, ensuring it knows to call check_url_safety before any URL access (WebFetch, Bash curl/wget, etc.).
• PostToolUse -- After every tool call, security rules are re-injected so Claude knows to call check_text_safety on any external content received.

Skills (Claude-driven) -- Four skills that teach Claude when and how to call Agent Looker's MCP tools:

| Skill | Trigger | Purpose | |-------|---------|---------| | check-url-safety | Before accessing any URL (curl, wget, git clone, etc.) | Claude calls check_url_safety before every URL access | | check-text-safety | When processing external text from any source | Claude calls check_text_safety on all received external content | | report-risk-url | Proactively, when a suspicious URL is discovered | Phishing, malware, scam, suspicious redirects | | report-risk-text | Proactively, when suspicious text is discovered | Prompt injection, jailbreak, data leaks |

> Note: Unlike the full Claude Code plugin, hooks in this Cowork edition do not make direct API calls — they inject security rules that guide Claude to use the MCP skills. All actual threat detection goes through the skills.

How protection works

WebFetch(url)
      |
      v
PreToolUse hook: injects security rules into Claude's context
      |
      Claude calls check_url_safety (MCP skill)
      |
      +-- UNSAFE --> Claude blocks the fetch and informs the user
      |
      +-- SAFE   --> WebFetch executes
                        |
                        v
               PostToolUse hook: injects security rules into Claude's context
                        |
                        Claude calls check_text_safety (MCP skill)
                        |
                        +-- BLOCK/FLAG --> Claude warns the user
                        +-- ALLOW     --> pass through

A safe URL can still serve malicious content. URL checks and content checks are two independent layers.

Requirements

• Claude Cowork with MCP support
• Agent Looker MCP server configured in your Cowork workspace
• An Agent Looker account (sign up at the dashboard)

Installation

1. Configure the Agent Looker MCP server

Add the Agent Looker MCP server to your Claude Cowork workspace settings. The MCP server provides the check_url_safety, check_text_safety, report_risk_url, and report_risk_text tools that skills call into.

Refer to your Agent Looker dashboard for the MCP server URL and authentication token.

2. Load hooks

Copy hooks/hooks.json into your Claude Cowork hooks configuration, or reference it from your workspace's hook settings. The hooks inject Agent Looker security rules into Claude's context before and after WebFetch and WebSearch calls.

3. Load skills

Copy the skills/ directory into your Claude Cowork skills directory. Each subdirectory contains a SKILL.md that teaches Claude when and how to call the corresponding MCP tool.

4. Restart Claude Cowork

Restart your Claude Cowork session to activate the hooks and skills.

Project structure

hooks/
  hooks.json           # PreToolUse / PostToolUse hook definitions
                       # (injects security rules via additionalContext)
skills/
  check-url-safety/    # Skill: check URLs before access
  check-text-safety/   # Skill: check text content safety
  report-risk-url/     # Skill: report suspicious URLs
  report-risk-text/    # Skill: report suspicious text

Difference from the full Claude Code plugin

| Feature | Claude Code plugin | Claude Cowork plugin | |---------|-------------------|---------------------| | PreToolUse URL blocking | Calls API directly, blocks before fetch | Injects rules; Claude calls MCP skill | | PostToolUse content scan | Calls API directly, warns via context | Injects rules; Claude calls MCP skill | | Setup script | Yes (bin/setup.mjs) | No | | Authentication | ~/.agent-looker.cfg | Via MCP server configuration | | Node.js required | Yes (for hook scripts) | No |

License

GPL-3.0 -- see LICENSE for details.