MCP Server

SafeLLM includes a built-in MCP stdio server with a plugin architecture. It exposes core security capabilities as tools that can be called from MCP clients.

Why MCP in SafeLLM

• Reuse SafeLLM detection stack outside HTTP gateway path.
• Keep one policy engine for APISIX traffic and agent workflows.
• Extend capabilities via plugin-style MCP tools.

Built-in tools

• safellm.guard_decide
  • Runs full input pipeline (cache -> keywords -> PII -> AI guard when enabled).
  • Inputs: prompt, uri, optional request_id.
• safellm.pii_scan
  • Runs PII layer only.
  • Inputs: text, optional request_id.
• safellm.dlp_scan
  • Runs output DLP scanner.
  • Inputs: text, optional mode (block|anonymize|log).

Run server (OSS)

From safellm-oss:

source .venv/bin/activate
python -m sidecar.mcp

Run server (Enterprise)

From safellm-enterprise/safellm:

source .venv/bin/activate
python -m sidecar.mcp

JSON-RPC methods

• initialize
• ping
• tools/list
• tools/call

Quick smoke test

printf '%s\n%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}' \
  | python -m sidecar.mcp

Plugin architecture details

Core plugin API lives in:

• sidecar/mcp/base.py
• sidecar/mcp/registry.py

Built-in plugin implementations:

• sidecar/mcp/plugins/security.py

Server bootstrap and stdio runtime:

• sidecar/mcp/server.py
• sidecar/mcp/__main__.py

To add custom tools, implement MCPToolPlugin, register it in MCPToolRegistry, and expose it during server bootstrap.

Reference deployment

• Docs: /deployments/mcp-reference/
• Repo folder: safellm-oss/examples/mcp-reference/