Creativity · MCP — server

MCP Modal Server

The MCP Modal Server integrates Modal's serverless Python runtime. MCP clients can list deployed apps, invoke functions, fetch logs, and manage scheduled functions or webhooks. It's a popular pairing for AI engineers running LLM inference, vector search, or fine-tuning jobs on Modal's GPU fleet.

MCP facts

Kind: server
Ecosystem: anthropic-mcp
Language: Python (modal SDK)
Transports: stdio

Capabilities

Tools: list_apps, invoke_function, list_logs, list_scheduled_functions, list_containers
Resources: modal://app/{name}, modal://function/{app}/{fn}
Auth: MODAL_TOKEN_ID + MODAL_TOKEN_SECRET

Install

pipx install mcp-server-modal

Configuration

{
  "mcpServers": {
    "modal": {
      "command": "mcp-server-modal",
      "env": {
        "MODAL_TOKEN_ID": "${MODAL_TOKEN_ID}",
        "MODAL_TOKEN_SECRET": "${MODAL_TOKEN_SECRET}"
      }
    }
  }
}

Frequently asked questions

Can the LLM deploy new Modal apps?

Deployment still happens via `modal deploy` from a repo. MCP focuses on runtime operations: invoking, inspecting, scheduling.

Is GPU cost a concern?

Yes — Modal bills per second of GPU. Gate invoke_function calls behind explicit approval for GPU-hungry functions, or wrap them in a small spend budget.

How does it fit in agentic workflows?

A Modal function can host a domain-specific tool (OCR, classifier, RAG step). MCP exposes it to Claude as a callable tool without needing Claude to manage the infrastructure.

Sources

Modal — serverless Python — accessed 2026-04-20
Model Context Protocol — accessed 2026-04-20
MCP servers repo — accessed 2026-04-20