cipherwake-mcp

MCP server exposing Cipherwake's deploy-gate + scan + preview-diff + suggest-fix tools to any Model Context Protocol-compatible AI coder.

Once installed, your AI agent can call these tools directly as part of its reasoning loop — no npx pqcheck step, no manual scanning. Customer asks "is this safe to deploy?" → Claude calls cipherwake_before_deploy_check → result lands in the chat reply.

Tools

| Tool | When the agent should call it | |---|---| | cipherwake_before_deploy_check | After finishing a change, before telling the user a deploy is ready. Returns pass / review / block. | | cipherwake_preview_diff | After deploying to a preview URL, before merging. Catches new third-party scripts, header regressions, soft-404s, broken protected paths. | | cipherwake_scan_domain | When the user asks "how secure is X?" about any HTTPS domain. Returns DBR grade + top findings. | | cipherwake_explain_finding | When the user asks "what does finding headers.csp.missing mean?" | | cipherwake_suggest_fix | When the user asks "what should I fix first?" Returns the highest-impact, customer-actionable fix. |

Every tool response includes a human_summary and required_user_summary field so the agent surfaces the result in its final reply instead of hiding the check.

Install

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%/Claude/claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "cipherwake": {
      "command": "npx",
      "args": ["-y", "cipherwake-mcp"]
    }
  }
}

Restart Claude Desktop. The 5 tools above appear in Claude's tool list.

Claude Code (CLI)

Add to your ~/.claude/settings.json:

{
  "mcpServers": {
    "cipherwake": {
      "command": "npx",
      "args": ["-y", "cipherwake-mcp"]
    }
  }
}

Cursor

Edit ~/.cursor/mcp.json:

{
  "mcpServers": {
    "cipherwake": {
      "command": "npx",
      "args": ["-y", "cipherwake-mcp"]
    }
  }
}

Restart Cursor. Tools appear in Cursor's MCP-mode tool list.

Cline (VS Code extension)

Open the Cline panel → MCP Servers → Edit cline_mcp_settings.json:

{
  "mcpServers": {
    "cipherwake": {
      "command": "npx",
      "args": ["-y", "cipherwake-mcp"],
      "disabled": false
    }
  }
}

Windsurf / Continue / other MCP clients

Same shape — point the client's MCP config at:

command: npx
args: ["-y", "cipherwake-mcp"]

Configuration

cipherwake-mcp reads two environment variables (both optional):

• CIPHERWAKE_API_BASE — override the API base URL (default: https://cipherwake.io)
• CIPHERWAKE_API_KEY — authenticated API key (qpk_...) for the per-account quota path. Without it, you fall under the per-IP free tier (120 scans/hour shared with npx pqcheck).

How the agent uses it

Anthropic's MCP doesn't force an agent to call a tool, but it strongly nudges. Cipherwake's tool descriptions are written to match agent lifecycle moments — cipherwake_before_deploy_check literally says "Use this AFTER finishing a code change and BEFORE telling the user a deploy is ready." That's the kind of phrasing agents respond to.

Combined with Cipherwake's AI Coder Protocol (installed via npx pqcheck protocol install), the agent has both:

1. A rule telling it to gate every deploy on ship_decision
2. A tool to actually run the check

The protocol rule + MCP tool together make the agent's behavior reliable across Claude Desktop, Claude Code, Cursor, Cline, and any future MCP client.

Route assertions — declare your private routes

If your repo has a .cipherwake.json at the root (see the methodology), Cipherwake asserts your declared private routes are still gated on every check the MCP tool runs. Drop a config like:

{
  "routeAssertions": {
    "assertions": [
      { "path": "/api/admin/users", "expect": "protected", "why": "User mgmt API" },
      { "path": "/api/internal/cron", "expect": "protected", "why": "Internal cron handler" }
    ]
  }
}

The agent sees ship_decision_assertions=block and the per-assertion CIPHERWAKE_ROUTE_ASSERTIONS block in the tool response. A critical failure (declared protected route now returns 200) auto-blocks via ship_decision=block — the agent stops and surfaces the change to the user, same protocol as drift-based blocking. No credentials required: Cipherwake never crawls behind your login. See why not for the design decision.

Brand visibility

Every MCP response from cipherwake-mcp includes a markdown summary like:

🛡️ **Cipherwake** · ⚠️ **REVIEW** — found 2 trust-surface changes since baseline:
- [HIGH] New third-party script: widget.intercom.io
- [MEDIUM] Strict-Transport-Security weakened: max-age=31536000 → max-age=3600

Agents are expected to paste this verbatim into their final reply. The brand stays visible to the user — Cipherwake doesn't run silently in the background.

Free vs paid

The MCP server is free to install + use. Underlying API quotas:

• Free (anonymous, per-IP): 120 scans/hour shared with npx pqcheck. Fine for ad-hoc checks; can be exhausted by heavy CI use.
• Cipherwake Founder Pro ($19.99/mo launch pricing, locked while subscription active): 5,000 calls/month + custom CI thresholds + Slack/webhook delivery + saved baselines. Set CIPHERWAKE_API_KEY to use this path.

See cipherwake.io/pricing.

Source + feedback

• Source: github.com/cipherwakelabs/pqcheck/tree/main/mcp-server
• Bug reports + feedback: cipherwake.io/feedback
• Cipherwake methodology (every probe documented): cipherwake.io/methodology

License

MIT