I'm excited to share the first of many upcoming open source projects: snowfakery-mcp, an MCP Server for Salesforce.org's Snowfakery tool.
The Problem with Test Data
Snowfakery has been open source for years. It's a powerful tool for declarative synthesis of complex relational data - used inside Salesforce and throughout the community. But authoring good Snowfakery recipes has always been a challenge.
I've tried multiple times over the years to see what LLMs could learn about Snowfakery. The tool has great docs and examples, but those tend to just flood context windows without ever producing a usable recipe.
The MCP Approach
Instead of cramming documentation into prompts, snowfakery-mcp gives Claude tools to learn the language interactively:
- Explore the Snowfakery syntax
- Understand recipe structure
- Validate recipes before running them
- Generate data that respects relational constraints
The result? Single-shot recipe generation that actually works.
See It In Action
The demo video shows a single request session with Claude Opus 4.5 using snowfakery-mcp to learn about the language and write a valid recipe.
No prompt engineering. No few-shot examples. Just tools that let the AI explore and learn.
Try It
Claude Desktop: Grab the MCPB Desktop Extension file to try it out.
MCP Registry: Find it as io.github.composable-delivery/snowfakery-mcp
Generated Recipe: Check out the recipe Claude created in that single conversation.
Why This Matters
This is what I mean by captured context. Snowfakery's knowledge was always there - in docs, examples, source code. But it wasn't accessible to AI in a useful way.
MCP servers are a form of context capture. They package domain expertise into tools that AI can actually use. The knowledge doesn't disappear into a context window - it becomes capability.
More MCP servers coming soon. This is just the beginning.
snowfakery-mcp is open source under MIT license. Contributions welcome.