# Escape for Claude Connect Escape to Claude Desktop, Claude.ai, and Claude Code through OAuth 2.1 without copying an API key. Once connected, you can use plain-English prompts to explore ASM assets, run AI pentests, triage findings, and prepare remediation guidance. Access follows your Escape user permissions. The same setup works for any MCP-compliant assistant. See [Other MCP clients](<#other-mcp-clients>) below for Cursor, ChatGPT desktop, Continue.dev, Zed, Windsurf, and Codeium. ## Why Escape for Claude AppSec teams are already living inside their AI assistants. Pulling Escape findings into Claude lets you: - Ask security questions in your team's own language ("show me unauthenticated endpoints exposed this week" instead of clicking through filter dropdowns). - Update specific issues' statuses after reviewing their evidence. - Hand developers a clean remediation in seconds (Claude reads Escape's evidence and reproduction payload, drafts the fix description, and tells you why it's a risk). - Generate custom reports on demand without writing a public API script. OAuth 2.1 lets you connect without copying an API key. The assistant caches your API key as its bearer token. The token response advertises a one-year lifetime; rotating or revoking the key invalidates cached tokens across your assistants on the next call. Treat the cached token like an API key. ## Setup in Claude The whole flow is five clicks. You'll need an Escape account; OAuth grants Claude access using the same API key your CLI uses, without requiring you to copy or paste it. ### Step 1: Open the Connectors Panel In Claude Desktop or Claude.ai, open **Settings → Customize → Connectors**, click the **\+** at the top of the connector list, and pick **Add custom connector**. ### Step 2: Add the Escape MCP URL Fill in the dialog and click **Add**: - **Name**: `Escape for Claude` (or anything you like; this is the local label). - **URL**: `https://mcp.escape.tech/mcp` - Leave **Advanced settings** untouched. ### Step 3: Click Connect The connector now appears under **Not connected** with a **Connect** button. Click it to start the OAuth handshake. ### Step 4: Approve in Your Browser Claude opens your browser to `https://app.escape.tech/oauth/mcp/authorize`. If you're already logged into the Escape dashboard the consent card appears immediately; otherwise sign in first. Click **Allow** and you'll see this confirmation: ### Step 5: Tune the Per-Tool Permissions Back in Claude, the connector flips to **Connected** and the full Escape tool catalog appears. Use the per-tool toggles on the right to set permissions: - **Always allow** for read-only tools (asset listing, knowledge questions, capability discovery). - **Needs approval** for write/delete tools (creating scan profiles, updating issues, starting pentests). Claude defaults to safe values, but you can tighten or loosen as needed. ## Use Cases The whole point of MCP is that Claude figures out which tool to call. A few prompts that work today, mapped to the AppSec workflow: - **Crown-jewel discovery (ASM)**: *"Show me every external asset Escape discovered this week that's unauthenticated and exposed on the internet. Group by hostname."* - **Conversational triage**: *"Pull the highest-severity issues from the last DAST scan on `api.example.com`. Group by category and help me review individual issues before updating their status."* - **AI Pentest Assessment**: *"Start an AI Pentest on `api.example.com` with our authentication preset. When it's done, summarize the business-logic findings and rank them by exploitability."* - **Remediation handoff**: *"For issue ``, draft a fix description aimed at the developer who owns the endpoint. Include the reproduction payload and explain why it's a risk."* - **Custom report generation**: *"Generate a PDF report for all SSRF findings on this API surface from this quarter."* - **CI/CD scan configuration**: *"Draft a YAML scan profile for our staging GraphQL endpoint with auth, then explain what each parameter does."* - **Knowledge Q&A**: *"What's the difference between Escape's BOLA detection and a generic IDOR check?"* These are seeds. Claude chooses the tool chain, you read the result, and you keep refining in plain English. > [!TIP] > > **Tell us what you build** > > Share your workflows and feature requests on your dedicated support channel or at `support@escape.tech` to help shape MCP improvements. ## Other MCP Clients The same `https://mcp.escape.tech/mcp` URL works for every MCP-compliant assistant, including ones behind your own MCP gateway or proxy: | Client | Setup | | --- | --- | | Cursor | `~/.cursor/mcp.json` with the URL only; OAuth on first use. | | ChatGPT Desktop / Codex | Add as a custom MCP server with the URL only; OAuth on first use. | | Continue.dev | Add to your Continue config with the URL only; OAuth on first use. | | Zed | Configure as an MCP server with the URL only; OAuth on first use. | | Windsurf / Codeium | Add via the MCP settings panel with the URL only; OAuth on first use. | For headless contexts (CI scripts, custom integrations, your own MCP client), keep using `Authorization: Key ` or `X-ESCAPE-API-KEY: `. Both flows use your Escape user permissions. Full setup details for each client live in the [IDE Integration Guide](). ## Troubleshooting ### The Connect Button Never Appears Claude has cached an OAuth bearer token from a previous session. To force a fresh handshake: 1. Click the **`…`** menu on the connector row and pick **Disconnect**. 2. Click **Connect** again. As a server-side reset, rotating your API key from the [Escape dashboard]() invalidates every cached token (across every assistant) on the next call. ### `redirect_uri is not allowed` During the OAuth Handshake Callback hosts don't need to be registered with Escape, so this points at the shape of the URI rather than the host. Cleartext `http://` is only accepted on loopback (`127.0.0.1`, `localhost`, `[::1]`); every other callback must be `https://` or a native custom URI scheme such as `cursor://`. If your client sends something else, [contact support](). ### Claude Shows Tools but Tool Calls Fail with `401` Your underlying API key was rotated or revoked. Disconnect and reconnect the connector to redo the OAuth handshake against your new key. ### I Don't Want to Use OAuth The legacy header form keeps working. Configure the connector with: ```http Authorization: Key ``` Or, equivalently: ```http X-ESCAPE-API-KEY: ``` See [IDE Integration Guide → static API key flow]() for full configuration examples. ## Related Documentation - [Escape MCP server overview]() - [IDE Integration Guide]() - [Escape Copilot]() - [Model Context Protocol Specification]() - [MCP 2025-06-18 Authorization Spec]() ## Questions? Have a question? Reach out on your dedicated support channel (Slack, Microsoft Teams, or whichever channel we've set up with your team), or email us at `support@escape.tech`.