MCP Integration

The ki-mcp binary speaks Model Context Protocol (MCP) over stdio — no server to run, no daemon, no config beyond a JSON snippet. Once registered, every AI tool in your workspace routes prompts through Ki!‘s local privacy engine before they reach any cloud model.

Tools Exposed

Tool	What it does
mask_text	Scrubs PII. Returns masked text + pii_count + regulatory_flags
reveal_tokens	Swaps [TYPE_xxxxxxxx] tokens back to originals within the same session

---

VS Code (v1.99+)

Install the Ki! Privacy Proxy extension from the Marketplace — it auto-registers the MCP server on startup.

Or add manually to settings.json:

{
  "mcp.servers": {
    "ki": {
      "command": "ki-mcp",
      "args": []
    }
  }
}

---

Cursor

Open Cursor Settings → Features → MCP Servers → Add, or edit ~/.cursor/mcp.json:

{
  "mcpServers": {
    "ki": { "command": "ki-mcp", "args": [] }
  }
}

---

Windsurf (Codeium)

Add to ~/.codeium/windsurf/mcp_config.json (create if missing):

{
  "mcpServers": {
    "ki": {
      "command": "ki-mcp",
      "args": []
    }
  }
}

Restart Windsurf. The tools appear automatically in the Cascade panel.

---

Claude Code (CLI)

claude mcp add ki -- ki-mcp

Or add manually to ~/.claude.json:

{
  "mcpServers": {
    "ki": {
      "command": "ki-mcp",
      "args": [],
      "type": "stdio"
    }
  }
}

---

Claude Desktop

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

{
  "mcpServers": {
    "ki": {
      "command": "ki-mcp",
      "args": []
    }
  }
}

---

Amazon Q Developer

Add to ~/.aws/amazonq/mcp.json:

{
  "mcpServers": {
    "ki": {
      "command": "ki-mcp",
      "args": [],
      "timeout": 10000
    }
  }
}

---

OpenAI Codex CLI

Add to ~/.codex/config.yaml:

mcp_servers:
  - name: ki
    command: ki-mcp
    args: []

---

Finding the ki-mcp Binary

After installing Ki!, the binary is bundled at:

Platform	Path
macOS	/Applications/Ki!.app/Contents/MacOS/ki-mcp
Linux	~/.local/bin/ki-mcp or /usr/local/bin/ki-mcp
Windows	%LOCALAPPDATA%\Ki!\ki-mcp.exe

To make it globally available:

# macOS
ln -s /Applications/Ki!.app/Contents/MacOS/ki-mcp /usr/local/bin/ki-mcp

# Linux
sudo cp ki-mcp /usr/local/bin/ki-mcp && sudo chmod +x /usr/local/bin/ki-mcp

---

Usage Example

Once connected, ask your assistant:

Mask this before sending: "Contact Alice Johnson at alice@acme.com, card 4111 1111 1111 1111"

The assistant calls mask_text. Only the masked version reaches the cloud:

Contact [PERSON_a1b2c3d4] at [EMAIL_e5f6a7b8], card [CREDIT_CARD_c9d0e1f2]

To restore the originals in the response, call reveal_tokens.

---

Configuration

~/.config/ki/mcp_config.json (created automatically on first run):

{
  "ocultar_url": "http://127.0.0.1:4141",
  "rate_limit_rpm": 60
}

Key	Default	Notes
ocultar_url	auto-detected	Override the Ki! sidecar address
rate_limit_rpm	60	Max tool calls per minute

---

MCP Registry Listings

Ki! MCP is listed on the major MCP registries:

• glama.ai — ki-mcp
• smithery.ai — @getki/ki-mcp
• mcp.run — ki-mcp

---

Python MCP Extensions (Ocultar)

For teams using Python-based LLM pipelines, the Ocultar MCP Extensions (pip install ocultar-mcp) expose the same mask_text and reveal_tokens tools inside LangChain, LlamaIndex, and any openai-SDK workflow — without requiring the ki-mcp binary.

pip install ocultar-mcp

from ocultar_mcp import OcultarMCPServer

server = OcultarMCPServer(ocultar_url="http://127.0.0.1:4141")
server.start()  # exposes MCP over stdio or SSE

The Python package is open source and free. It connects to any running Ocultar Refinery instance — the Ki! Desktop sidecar, or a self-hosted Ocultar deployment.

For organisation-wide enforcement (not just individual developer workstations), see Sombra Gateway, which provides a single drop-in endpoint covering every service and team.