WellSet

Inspiration

Laboratory experiments are often arranged in 96-well plates: grids of small wells where each position can hold one experiment. Position can influence results, while scientists also possess local knowledge that generic automation cannot predict. A control may need to remain fixed, or a well may be physically unavailable.

Repairing the entire layout manually is tedious. Letting an agent rearrange it without respecting those constraints is unsafe.

WellSet explores a better pattern: Local judgment. Global repair.

What it does

WellSet gives the scientist and ChatGPT one shared, live experimental layout.

The scientist expresses contextual judgment directly on the plate:

• lock controls that must not move;
• block wells that cannot be used;
• inspect visible validation violations.

ChatGPT then invokes a semantic WebMCP Site Tool to reflow only the unlocked experiments. The application validates the result, preserves human constraints and produces a compact operation receipt.

In the demonstrated hero flow:

• controls B2 and B3 are locked;
• occupied wells B6 and B7 are blocked;
• two violations become visible;
• ChatGPT selects wellset_reflow_unlocked from an ordinary-language request;
• 24 experiment placements move;
• both locked controls remain unchanged;
• both blocked wells are avoided;
• validation falls from two violations to zero.

Why WebMCP

Without WebMCP, an agent may need to infer meaning from pixels, reverse-engineer the DOM or perform fragile coordinate-based clicks across a dense grid.

WellSet instead exposes three stable, bounded, imperative tools:

• wellset_inspect_plate
• wellset_generate_layout
• wellset_reflow_unlocked

These tools give ChatGPT meaningful operations rather than low-level interface mechanics. The person and agent operate on the same underlying state, so human edits immediately become constraints on the agent’s next action.

This is the core WebMCP advantage demonstrated by WellSet: the website can explicitly communicate what an agent may do, while retaining final domain validation inside the application.

How we built it

WellSet is a client-side web application with no backend, account, remote content or model API. The interface and all three WebMCP tools share one deterministic state store.

The implementation uses:

• static top-level WebMCP tool registration;
• closed input schemas and runtime validation;
• bounded structured outputs;
• atomic, idempotent mutations;
• fail-closed constraint enforcement;
• application-owned validation;
• auditable operation receipts.

The reflow tool cannot move locked assignments or place experiments in blocked wells. A proposed mutation is committed only after the complete resulting plate passes domain validation.

Challenges

The hardest challenge was not merely registering tools. It was ensuring that the visible interface, tool state, validation result and receipt could never disagree.

We also had to make a specialised scientific interaction understandable to non-scientists, keep the tool surface small enough for reliable agent selection and make the hero flow deterministic enough for repeated judging.

Accomplishments

The frozen release passed:

• 136/136 automated tests;
• five out of five fresh ChatGPT Site Tools trials;
• discovery of all three tools in every trial;
• correct operation selection from natural language in every trial;
• preservation of locked controls in every trial;
• reduction from two violations to zero in every trial;
• exact agreement between receipts and the visible UI in every trial.

Median completion time across the five fresh trials was 38 seconds.

What we learned

Agent-native interfaces become more trustworthy when humans express judgment locally, agents receive semantic capabilities instead of inferred clicks, and applications remain responsible for enforcing invariants.

A small, precise tool surface can be more powerful than a large collection of overlapping actions.

What’s next

WellSet is currently a synthetic technical demonstration, not validated laboratory or clinical guidance.

A production pathway would require collaboration with laboratory practitioners, scientifically validated placement policies, persistent audit records, broader accessibility testing and evaluation across real experimental workflows.

The broader pattern extends beyond laboratory plates: people define local constraints on a shared artifact, while agents perform bounded global transformations without erasing human judgment.

Built With

• chatgpt
• css3
• html5
• javascript
• site
• tools
• vercel
• webmcp