> ## Documentation Index
> Fetch the complete documentation index at: https://docs.breachlens.app/llms.txt
> Use this file to discover all available pages before exploring further.

# MCP servers

> Register a Model Context Protocol server and run BreachLens's 10-probe MCP suite — tool poisoning, prompt injection, rug-pulls, token theft, and more.

Model Context Protocol (MCP) servers hand tools to AI agents — and a malicious or misconfigured one is a direct path into whatever the agent can touch. BreachLens connects to an MCP server, enumerates its tools, and runs its **10-probe MCP suite** — a set of BreachLens MCP threat classes informed by public MCP-threat research and OWASP-LLM concepts: tool poisoning, prompt injection, rug-pulls, tool shadowing, indirect injection, token theft, malicious code execution, excessive permission scope, remote access, and a multi-vector check.

## Register a server

Open **MCP Servers → Add server** and describe how to reach it:

* `Name` · `Description` (optional)
* `Transport` — how the server speaks (see [Transports](#transports) below)
* `Endpoint` — the command or URL for the chosen transport.
* Optionally link it to an Application, Repository, or Container for cross-tier context.

### Transports

| Transport     | Endpoint value                                 |
| ------------- | ---------------------------------------------- |
| **STDIO**     | the process command to launch (a local server) |
| **HTTP**      | the server URL                                 |
| **SSE**       | the server URL                                 |
| **WebSocket** | a `wss://` URL                                 |

<Note>
  **Authenticated remote servers are coming.** Today you can point BreachLens at a local (STDIO) server or an unauthenticated HTTP / SSE / WebSocket endpoint; per-server bearer-token auth isn't configurable yet. If you need to test a token-protected MCP server, [tell us](mailto:sales@breachlens.app?subject=BreachLens%20MCP%20auth).
</Note>

## Run a scan

<Steps>
  <Step title="Open the server">
    Go to **MCP Servers** and click the registered server.
  </Step>

  <Step title="Run the scan">
    Click **Run scan**. BreachLens opens the transport, completes the MCP handshake, enumerates the exposed tools, and runs the probe suite against them. If the handshake fails, you get a clear connectivity finding instead of a silent miss.
  </Step>

  <Step title="Review findings">
    Each probe can emit zero, one, or more findings, each classified by its BreachLens MCP threat class.
  </Step>
</Steps>

<Note>
  Some probes (prompt injection, tool poisoning, indirect injection) are **LLM-judged** — they reuse the same judge model as [AI-agent scans](/ai-identity/ai-agents). With no AI provider configured they simply skip; the heuristic probes still run, so a scan always produces a result.
</Note>

## Next steps

<CardGroup cols={2}>
  <Card title="Red-team an AI agent" icon="robot" href="/ai-identity/ai-agents">
    Probe the agents that consume these MCP servers for prompt injection and more.
  </Card>

  <Card title="Fit it into your pipeline" icon="diagram-project" href="/pipeline/cicd">
    Where AI and MCP testing sit alongside the rest of your security pipeline.
  </Card>
</CardGroup>