Author carefully. Validate with the engine.
Guidance for AI agents authoring LazyAnalysis graphs through the MCP server, without inventing node settings or pin names.
1. Describe the task
Identify inputs, units, expected outputs, sampling assumptions, and required visualizations. Distinguish a conceptual workflow from a graph ready to execute.
2. Use the MCP server
The LazyAnalysis MCP server gives an agent the node catalog exported from the application, builds graph files from a short spec, and checks them the way the application loads files. Prefer it to writing graph JSON by hand: it copies each node’s real saved fields and assigns consistent identifiers.
3. Wire compatible pins
Name pins by their exported identifier, their title, or their position (in, out, in 2). Links go from an output to an input of the same type, and an input takes one link. The server reports a wrong pin name with the node’s actual pins.
4. Validate in the application
The server’s checks catch files that would not load. Whether a graph computes what you intended is confirmed by opening it in LazyAnalysis, running it with controlled inputs, and comparing outputs with expected values.
5. Add an interface where supported
Add dashboard cards for nodes whose UI supports them, and notes that explain the graph. Running graphs and reading results from the agent come with the hosted browser workspace. llms.txt is navigation guidance, not a training mechanism.