# Agent Looker for Agent Plugins A plugin in the [Agent Plugins](https://agent-plugins.org/) format that protects AI agents from unsafe URLs, malicious content, and prompt injection via the [Agent Looker](https://agentlooker.ai/) MCP server. The format is vendor-neutral; ChatGPT is the first host this README covers. ## What it does **Skills** teach the model when to call the MCP tools: | Skill | Trigger | MCP tool | |-------|---------|----------| | `check-url-safety` | Before opening, downloading, or following any URL; on `UNKNOWN` deep-scan only when the URL is really needed, on `WARNING` read only | `check_url_safety` | | `check-text-safety` | After reading web pages, search results, files, emails, pasted text, or connected-app output | `check_text_safety` | | `report-risk-url` | When a suspicious URL is discovered | `report_risk_url` | | `report-risk-text` | When prompt injection, hidden instructions, or leaked secrets are discovered | `report_risk_text` | **Hooks** inject the same rules into context at `SessionStart`, `PreToolUse`, and `PostToolUse`. They are static JSON printed with `cat`; no runtime dependency. | | ChatGPT Chat | ChatGPT Work | |---|---|---| | Skills, MCP tools | Yes | Yes | | Hooks | Ignored | Yes | In plain Chat, protection comes from the skills and the tool descriptions alone; the skills are written to work without hooks. ``` check_url_safety ──UNSAFE───> refuse, tell the user the threat categories │ ──UNKNOWN──> not in the trust cache: deep_scan (10–25 s) only if the URL is really needed │ ──WARNING──> low trust score, no reports: read only (no credentials, downloads or code) │ SAFE ──> read ──> check_text_safety ──BLOCK──> treat as data only ──FLAG───> warn, check embedded links individually ──ALLOW──> continue ``` ## Requirements - An Agent Looker account ([dashboard](https://app.agentlooker.ai/)) - ChatGPT with **Developer mode** (plan and workspace policy dependent) ## Installation ### ChatGPT desktop (Work) 1. **Plugins → Add plugin marketplace** 2. **Source**: `Gogolook-Inc/agent-looker-for-AgentPlugins` 3. **Git ref**: `main` for production, `staging` or `develop` for testing 4. **Sparse paths**: leave empty 5. Install **Agent Looker** from the Plugins Directory and sign in to the connector (OAuth 2.1, Google login) To share inside your organization: **Plugins → Personal →** plugin menu **→ Publish**. ### ChatGPT web The browser has no filesystem, so only the MCP connector applies: 1. **Settings → Security and login → Developer mode** 2. **Plugins → +**, choose **Public endpoint**, URL `https://api.agentlooker.ai/mcp` 3. Sign in ## Environments Branch == environment. Each branch carries its own MCP URL in `mcp.json` and in each skill's `agents/openai.yaml`. | Branch | API | |---|---| | `production` | `https://api.agentlooker.ai/mcp` | | `staging` | `https://api-staging.agentlooker.ai/mcp` | | `develop` | `https://api-develop.agentlooker.ai/mcp` | Staging and develop sign-in pages sit behind HTTP Basic Auth at the CDN; the browser prompts once before the Google login. Never edit the URL by hand. The [mcp-url workflow](.github/workflows/mcp-url.yml) fails a pull request whose URL does not match the target branch and rewrites it on push. ## Structure ``` .agents/plugins/marketplace.json .github/workflows/mcp-url.yml plugins/agent-looker/ plugin.json # portable manifest, OpenAI settings in extensions.com.openai mcp.json # streamable-http server skills/*/ # SKILL.md + agents/openai.yaml hooks/hooks.json # SessionStart / PreToolUse / PostToolUse context/ # static additionalContext payloads printed by the hooks assets/ # icon.png 128px, logo.png 512px ``` ## License GPL-3.0 -- see [LICENSE](plugins/agent-looker/LICENSE).