Local LLM

AgentGuard CB and the MerchantGuard MCP server are model-agnostic. Drive them from a fully local LLM via Ollama, LM Studio, or any Qwen-compatible runtime. The evidence pipeline never invokes an LLM, so the local model is only used to orchestrate calls.

Ollama

# Pull a tool-capable model (example: Qwen 2.5 32B)
ollama pull qwen2.5:32b

# Run an MCP-capable client (Claude Code, Cline, Continue.dev)
# pointing it at the local Ollama endpoint and the MerchantGuard MCP server.

LM Studio

• Open LM Studio → Discover → search for a tool-calling model (Qwen 2.5, Llama 3.1 with tools, etc.).
• Start the LM Studio local server (default http://localhost:1234).
• Configure your MCP client (Cline, Continue.dev, custom agent) to use the LM Studio endpoint.
• Add the MerchantGuard MCP URL to the same client.

Qwen-direct

• Qwen models run cleanly under vLLM or transformers locally.
• Use any MCP-compatible client to bridge between your Qwen runtime and the MerchantGuard MCP.
• No special configuration on the MerchantGuard side; the MCP server is transport-agnostic.

What still phones home

• AgentGuard CB self-hosted: nothing. Zero telemetry by default.
• MerchantGuard MCP (Cloud Run): the MCP server itself is hosted by us. If you want a fully air-gapped MCP, fork our repo and run the MCP locally.
• Stripe submissions: handled by your own Stripe SDK on your infrastructure.