Agents & MCP

machinable ships an MCP server that lets an AI coding agent drive research with you: author interfaces, run experiments, and ask scientific questions. It stays true to machinable's code-all-the-way-down approach rather than bolting on an experiment-tracking schema.

The premise: because experiments, measurements, and even statistical questions are ordinary interface code, an agent that can read and write that code can run the whole research loop, and machinable's contracts tell it what to write next.

The loop

Is SGD faster than Adam on this problem?

1. Frame the question as a comparison of operands by a measurable quantity.
2. Scaffold an aggregate interface that sweeps the experiment. The agent writes it with write_source, and machinable returns any import error so it can self-correct.
3. Launch it. Runs are content-addressed, so this is incremental.
4. Ask an inference (outperforms, converges_faster, ...). If an operand doesn't expose the needed <quantity>(), the inference returns a NotImplementedError naming the accessor to write. The agent writes it and re-runs, now a cached no-op for unchanged operands.
5. Surface the verdict as a widget for you to validate.

Every step is content-addressed, searchable, and reproducible, with no tracking layer.

What the server provides

• Tools: authoring (get_module, resolve_config, write_source with import-error surfacing, search_interfaces, launch), inference (outperforms, beats_baseline, ranks, converges_faster, infer, quantity), and widget verification (read_widget_result).
• Resources: the context the agent writes against, attached without tool calls: guide://authoring, inferences://catalog, examples://{name}, module://{module}, widget-sdk://docs.
• Prompts: canned workflows (investigate, scaffold-aggregate, add-inference, scaffold-widget).

Read on

• Setup: install and launch the server, connect a client.
• The research workflow: the investigate loop in detail.
• Authoring for agents: the tools, resources, and prompts.
• Reference inferences: the inferences that ship with it.

Why inferences instead of a metrics schema is covered in Design notes → Inference design.