🚀Example

Connect an agent to BioBox and ask your first grounded question

This walks you from zero to your first grounded answer in a few minutes. This is an example question workflow. The types of questions you can ask depend on the data you have available in your knowledge graph and how you have configured your graph schema.

1. Connect

If you haven't already, connect Claude to BioBox and sign in. You'll know it worked when BioBox shows as a connected tool in your Claude client.

2. Ask your first question

In a Claude conversation, ask:

What are the top 10 gene targets for ALS?

Claude will call BioBox behind the scenes — typically resolving ALS to a disease node, finding the relevant reasoning system, prioritization report, and reading its ranked scores — then answer with the actual top targets and their scores. You may see it mention the tools it used (like search_entities, list_reports, and get_report_scores); that's expected, and it's how you know the answer is grounded in your graph rather than the model's memory.

A good answer looks like a ranked list with scores, naming the report it came from — something you could open in the BioBox app and verify.

3. Follow the thread

The value of an assistant is the follow-up. Try:

Why was the target SOD1 deprioritized a few months ago?

What other indications is SOD1 a suitable target for?

What cell populations are most important for ALS?

What are the curent targets of phase 1 recuriting ALS trials and where do they rank in my target prioritization models ?

Each of these maps to a BioBox tool (get_evidence, where_entity_ranks, get_entity_connections), and Claude chains them for you.

Tips for grounded answers

• Name the disease, gene, drug, or model you care about — Claude resolves names to graph nodes, so specifics help it pick the right one.

• Ask for the source. "Which report?" / "Show the evidence." keeps answers traceable back into BioBox.

• If a name is ambiguous (a symbol that matches several nodes), Claude may ask you to pick the right one — that's the disambiguation step, not an error.

• It won't change anything. The MCP is read-only, so explore freely.

Where to next

• Tool Reference — everything Claude can call, with parameters and example responses, if you want to understand exactly what's happening under the hood.