nwb-mcp-server: Skill for Claude Code

NWB MCP Analysis

Two-phase workflow: explore with SQL first, then incorporate findings into code if the user needs more complex analysis or re-runnable code.

---

Phase 0: Confirm The Active Source

The server may start with no dataset configured at all. Do not assume a preset root_dir or dandiset_id exists.

Recommended order before any schema or SQL work:

1. Call get_active_source
2. If the source is unset, choose one in chat:
   • use_local_source(root_dir="...")
   • use_dandiset_source(dandiset_id="000363")
3. If the user wants to go back to any startup preset, call reset_active_source

Important behavior:

• Source selection is scoped to the current chat session, not global across all clients. You can still change the active source multiple times within the same chat session if needed, accepting the cost of rescanning or reinitializing the new dataset.
• nwb_file_search_code_snippet depends on the active source and will fail until one is selected.
• If a dataset-dependent tool says no dataset is active, switch sources first instead of retrying the same query.

---

Phase 1: Explore

Recommended order:

1. get_active_source if you have not checked the source yet
2. get_tables to see what is available
3. get_nwb_paths to check how many files are in the dataset before diving in; the count shapes how you interpret aggregate results
4. get_table_schema on tables of interest
5. execute_query for targeted analysis

Avoid preview_table_values unless you need actual cell values. get_table_schema is faster and cheaper for schema discovery.

TimeSeries tables (those with timestamps and data columns) can be huge when concatenated across many files or sampled at high rates. Before querying them, warn the user and propose a strategy:

• slice to specific time windows
• subsample every Nth row for a representative overview

Ask which approach they prefer unless the context already makes the choice obvious (e.g. slicing is a natural fit for trial-aligned analysis).