MCP (Model Context Protocol) as first-class service type

[richardkiene]

Parent Issue

#19 - Multi-language runtime support via base images

Depends On

• Python runtime support via componentize-py #21 - Python runtime support
• JavaScript/TypeScript runtime support via jco #22 - JavaScript runtime support

Overview

Make MCP (Model Context Protocol) servers a first-class service type in Fabricks, with built-in understanding of the protocol, capability restrictions, and tooling.

Motivation

MCP servers are an ideal use case for Fabricks:

• Security critical: MCP servers have tool execution capabilities
• Trust problem: Users want to run third-party MCP servers safely
• WASM sandboxing: Perfect fit for capability-based security

Target User Experience

[info]
name = "my-mcp-server"
type = "mcp"  # New service type

[from]
image = "fabricks.dev/runtimes/python:3.12"

[source]
entrypoint = "server.py"

[capabilities.mcp]
tools = ["read_file", "search"]  # Allowed tool names
resources = ["file://*"]          # Allowed resource patterns
max_context_size = "100KB"        # Limit context window usage

Features

1. MCP-aware service type

New type = "mcp" that understands:

• JSON-RPC protocol over stdio or HTTP/SSE
• Tool definitions and invocations
• Resource access patterns
• Prompt/context handling

2. Capability restrictions for MCP

Fine-grained control over what an MCP server can do:

[capabilities.mcp]
# Which tools the server can expose
tools = ["read_file", "write_file", "search"]

# Resource URI patterns allowed
resources = [
  "file:///allowed/path/*",
  "https://api.example.com/*"
]

# Limits
max_tool_response_size = "1MB"
max_concurrent_tools = 5

3. MCP protocol proxy

The daemon can optionally proxy MCP traffic to:

• Audit tool calls
• Enforce capability restrictions
• Rate limit requests
• Log for compliance

4. MCP discovery

• fabricks mcp list - Show available MCP servers
• fabricks mcp inspect <service> - Show tools/resources a server provides

5. Integration with Claude/AI clients

Easy connection string generation:

fabricks mcp connect my-mcp-server
# Outputs configuration for Claude Desktop, etc.

Implementation Phases

1. Basic MCP type: Recognize type = "mcp", treat as HTTP/stdio service
2. MCP capability model: Add [capabilities.mcp] section
3. Protocol-aware proxy: Inspect and enforce MCP traffic
4. Discovery and tooling: CLI commands for MCP management

Example: Running a Third-Party MCP Server

# Clone someone's MCP server
git clone https://github.com/someone/cool-mcp-server
cd cool-mcp-server

# Add a Fabrickfile
fabricks init --type mcp --runtime python

# Build and run with restricted capabilities
fabricks build
fabricks run .

References

• Model Context Protocol Spec
• MCP Python SDK
• MCP TypeScript SDK

Acceptance Criteria

• type = "mcp" recognized as valid service type
• [capabilities.mcp] section parsed and validated
• MCP servers can be built and run
• Example MCP server in examples/ directory
• Documentation for running MCP servers in Fabricks