@blursec/mcp

Official Model Context Protocol server for @blursec/sdk — lets AI agents (Claude, GitHub Copilot, Cursor, custom agents) check credentials, session tokens, and hashes against live stealer-log threat intel.

TL;DR — one npx command turns Blursec into five agent tools. All hashing happens locally; only a 10-character k-anonymity prefix ever leaves the machine. Zero runtime dependencies — the protocol core is hand-rolled.

Install & run

npm i -g @blursec/mcp @blursec/sdk   # or run via npx, see below

The server reads its configuration from the environment:

| env var | purpose | |-----------------------|----------------------------------------------| | BLURSEC_API_KEY | API key (required) | | BLURSEC_ENVIRONMENT | production (default) / staging / local |

Never pass the API key as prompt text or argv — use the host's env block.

Claude Desktop / Claude Code

{
  "mcpServers": {
    "blursec": {
      "command": "npx",
      "args": ["-y", "@blursec/mcp"],
      "env": { "BLURSEC_API_KEY": "bsk_live_..." }
    }
  }
}

VS Code (GitHub Copilot)

.vscode/mcp.json:

{
  "servers": {
    "blursec": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@blursec/mcp"],
      "env": { "BLURSEC_API_KEY": "${input:blursec-api-key}" }
    }
  }
}

Tools

| Tool | Wraps | Use it for | |------|-------|------------| | blursec_check_credential | credentials.check | "Is this email+password leaked?" | | blursec_check_token | credentials.checkToken | Session cookie / bearer token triage | | blursec_check_hash | credentials.checkHash | Pre-hashed checks — most private, the raw secret never enters the conversation | | blursec_verify_session | sessions.verify | Incident response on live sessions (auto-detects JWT vs opaque) | | blursec_whoami | auth.whoami | Verify connectivity + key scope |

Deliberately not exposed: auth.rotate — it invalidates the live API key immediately, which is far too destructive to hand to an autonomous agent.

Security model

• K-anonymity preserved — hashing happens inside the server process via the SDK; only the 10-hex-char prefix is transmitted to the Blursec API.
• No secret echo — tool results never repeat the input credential/token, so secrets are not duplicated into the model context.
• Fail-open is explicit — results carry failedOpen: true when the database was unreachable. The server's initialize instructions tell the model to treat that as unknown, never as safe.
• Zero runtime dependencies — like the SDK itself, this package ships no third-party code. The MCP protocol subset (stdio framing, initialize, tools) is implemented in ~200 audited lines.

Embedding in-process

The protocol core is exported for hosts that want a custom transport:

import { BlursecMcpServer, createBlursecFromEnv } from "@blursec/mcp";

const server = new BlursecMcpServer(createBlursecFromEnv());

// Feed it newline-delimited JSON-RPC; undefined means "notification, no reply".
const reply = await server.handleLine(line);

License

MIT © Blursec