Skip to main content
This walks you from nothing to a security assessment on one of your agents, driving Enoki from Claude Code over MCP. You need an account in an Enoki workspace. Workspaces are set up with our team; if your team already uses Enoki, a workspace admin can invite you. Using another client? Codex, VS Code, Claude Desktop, claude.ai and ChatGPT sign in the same way, and Cursor connects with an API key. Connect your agent shows each; then follow the same steps.
1

Connect Claude Code

Add the Enoki MCP server:
Then run /mcp in a session and authenticate enoki. Enoki opens in your browser: sign in to app.enokilabs.ai if you aren’t, check what Claude Code asks to do, and click Allow. Keep everything it asks for ticked: the steps below add a target, start a run and re-test a fix. Approving takes a workspace owner or admin; if the page says it can’t let Claude Code in, see Troubleshooting. To use an API key instead, see Connect your agent.
2

Register your target

Ask your agent to register your endpoint as a target. It should call get_setup_guide for the config reference, then create_target, which fires one live probe and only saves the target if it succeeds. On a failure it returns a categorized reason (invalid_url / unreachable / auth_failed / parse_error / no_reply / not_an_endpoint / unknown) so you can adjust and retry; Troubleshooting says what each means. An agent that needs a key, token or password is added in the Enoki app instead, by a workspace owner or admin: open app.enokilabs.ai/inventory and choose Add integration.To test an MCP server, ask your agent to register it the same way, from its address and how it signs in. A server without sign-in is saved at once. For one that signs in, your agent gives you a link instead: open it signed in to Enoki, confirm the server’s host, and add its key on that page, never in the chat. Then tell your agent you’re done, and it checks with get_pending_target. See Connect your MCP server.Every target then shows in list_targets with its model_id.
3

Run an assessment

With the target’s model_id, call trigger_assessment to start a run. It returns a run_id at once and the run continues in the background: poll get_assessment_status with that run_id until the status leaves running (completed, error or cancelled). See what we test for. A trial workspace can register an MCP server, but assessing one isn’t part of the trial, so trigger_assessment refuses it.
4

Read and fix findings

Pull results with list_findings and get_finding (which includes the repro). Once you’ve fixed one, close_finding it as fixed and run validate_fix to re-run its attacks and confirm the fix holds. An MCP server’s findings can’t be re-tested yet: run a new assessment to confirm a fix.

Next steps

MCP tool reference

Every tool, grouped by area, with the scope each needs.

REST API

Drive the same flow from CI/CD or your own code.