Inspiration
What it does
How we built it
Challenges we ran into
Accomplishments that we're proud of
What we learned
What's next for Particle Atlas
Inspiration
Particle physics is full of relationships that are difficult to understand from a table or chat response alone: particles can share charge and spin while differing enormously in mass, form matter/antimatter pairs, or connect through different force networks.
Particle Atlas began as an attempt to make the symmetries of the Standard Model visible in three dimensions. The WebMCP Challenge created an opportunity to solve the next problem: allowing a learner to move directly from a natural-language question to an inspectable visual investigation without giving up control of the interface.
What it does
Particle Atlas turns particle-physics questions into shared visual investigations. A learner can ask a browser agent to examine a pattern, and the agent uses eleven structured WebMCP tools to query Particle Data Group-linked data, configure the live Three.js scene, compare particles, reveal force networks, and save the evidence into an editable notebook.
Each finding captures the exact scene that produced it. The learner can rotate or filter the visualization, revise the agent's explanation, inspect and undo tool calls, reopen any saved view, and share the complete investigation through one URL. The agent is not controlling a hidden copy of the app; both participants work in the same visible state.
Before WebMCP, an agent had to infer controls from pixels and repeatedly manipulate a dense 3D interface. It could answer a question in text, but it could not reliably build a durable, replayable visual explanation inside the application. WebMCP makes that collaboration structured, inspectable, and reversible.
Why WebMCP
Particle Atlas combines read tools, state-changing tools, and a durable collaborative artifact:
- The agent queries application data instead of relying on model memory.
- It selects the correct axes, particle families, and force network for the question.
- Every tool call updates the learner's visible workspace.
- Findings bind reasoning to an exact scene snapshot.
- Human edits and agent actions remain synchronized.
- Undo and shared URLs make the work inspectable and reproducible.
A learner no longer needs to translate a conceptual question into a sequence of filters, axes, comparisons, camera positions, and force toggles. The agent handles that mechanical translation while the learner explores, challenges, edits, and shares the result.
How we built it
The application registers eleven tools through document.modelContext.registerTool() in web/js/webmcp.js. Their handlers call the same command layer used by manual controls in web/js/main.js.
Read tools expose the particle catalog, scene state, and investigation. Mutating tools focus, compare, configure, highlight, show force networks, frame the notebook, and capture findings. Each mutating tool records a full undo snapshot. The notebook module serializes human and agent edits and stores non-recursive scene snapshots for replay.
The application is a static Three.js site with no account, API key, or backend requirement. It remains fully usable in manual mode when WebMCP is unavailable.
Challenges we ran into
- Keeping manual actions, agent actions, camera state, comparisons, notebook edits, and shared URLs synchronized.
- Making every mutating agent action exactly reversible.
- Saving scene evidence without recursively embedding the investigation notebook.
- Distinguishing measured particle properties from plotting coordinates for massless or constrained particles.
- Maintaining a coherent full-screen 3D experience across desktop, mobile, table, light, dark, and high-contrast modes.
Accomplishments that we're proud of
- Eleven working WebMCP tools with structured schemas and results.
- A genuinely shared human-agent notebook with replayable scene evidence.
- Full-state agent undo and inspectable input/output history.
- Reproducible scene and notebook URLs.
- PDG-linked provenance, uncertainty, and hypothetical-particle labeling.
- Accessible table representation, keyboard navigation, mobile controls, and reduced-motion support.
- A green public browser regression workflow plus native Chrome and ChatGPT in-app browser verification.
What we learned
Agent-native design is not just exposing every button as a tool. The most useful tools operate at the level of user intent: compare a family, reveal an interaction, or save a finding with its evidence. The interface also needs visible state, reversibility, and a durable artifact so the human can remain an active collaborator.
What's next for Particle Atlas
- Classroom investigation templates.
- Exportable notebook reports.
- Decay, coupling, and composite-particle datasets with the same provenance model.
- Shared investigations for other scientific domains.
Log in or sign up for Devpost to join the conversation.