Structural Evolution

AI and humans co-engineer physical assemblies through typed WebMCP tools, visible topology evidence, and human-only approval.

Comment

Structural Evolution

One-line Summary

A browser-native workbench where an AI agent and a human co-steer an editable physical assembly through typed WebMCP tools and visible, reviewable topology evidence.

Problem

Early physical design is fragmented: a hardware engineer, a 3D model, a solver, and an AI assistant often operate in separate contexts. Screenshot-driven assistance cannot reliably see the active assembly revision, protected volumes, component positions, or whether a prior result is stale. That makes an agent's suggestion hard to trust and hard for a human to supervise.

Solution

Structural Evolution keeps the agent and human in one browser workbench. WebMCP exposes bounded, typed operations over the live design state: inspect the exact context, select an approved typed assembly, validate its layout, generate a bounded topology candidate, inspect or move components, and review named load cases. Every agent action is reflected in the same visible viewport, including topology geometry, protected regions, stress and displacement fields, replay motion, action receipts, and the human-only decision boundary.

Impact

For small drone and robotics teams, this is a more inspectable way to explore a physical concept before committing to a heavier engineering stack. It makes the agent useful without asking the human to trust hidden state or automatic promotion. The impact claim is deliberately narrow: the project demonstrates a browser-native collaboration pattern for early structural exploration, not a replacement for validated engineering or manufacturing approval.

AI Use

The product uses WebMCP as the agent interface. The page registers typed tools through document.modelContext.registerTool; schemas constrain input, tool descriptions describe the exact live state, and tool responses return bounded facts. The agent can inspect and validate the active assembly, request a deterministic topology candidate, and display named structural-estimate cases while the human continues to orbit, inspect, and control the shared viewport. State-changing import work is staged for human review, and candidate acceptance remains human-only.

Codex Use

OpenAI Codex was used to implement and iterate on the application, its typed tool contracts, tests, documentation, and release checks. Codex is also one of the tested WebMCP clients for the browser interaction shown in the demo flow. ChatGPT was used as an additional AI tool during the project.

Key Features

  • Shared editable 3D workbench with reference FPV-drone and SE-6 cobot demo assemblies.
  • Ten typed WebMCP tools spanning context inspection, approved-assembly generation, layout validation, topology generation and comparison, component operations, simulation review, and the cobot thermal study.
  • Same-origin module-worker topology execution with a Rust/Wasm sparse-SIMP reference solver, per-load signed displacement vectors, structural estimate fields, and immutable action receipts.
  • Agent-driven review of hover, aggressive roll, pitch braking, and yaw burst with synchronized full-assembly and frame-only views.
  • Retained real-time replay geometry with explicitly labelled 1,000x deformation, mounted components following their attachment points, and live stress/displacement coloring.
  • Review panel that separates agent prediction, interactive output/evidence, current plan state, and human authority.
  • Protected volumes, named load cases, connected load paths, stale-plan marking after edits, and no automatic candidate promotion.

Architecture

The public release is a static Vite + React + TypeScript application rendered with Three.js. WebMCP tools are page-local document.modelContext.registerTool registrations with bounded JSON schemas and read-only/untrusted-content annotations. The dashboard detects WebGPU capability; its interactive topology path runs in a same-origin module worker and dynamically loads Rust/Wasm. This release has no backend, remote solver, account system, or durable project storage.

Testing

No credentials are required.

  1. Open https://webmcp-structural-evolution.vercel.app in a current desktop browser. No credentials are required.
  2. In the ChatGPT/Codex in-app browser with WebMCP enabled, invoke the prompts in the demo script.
  3. Verify that the agent selects and validates the reference drone, generates a topology candidate, and cycles through hover, aggressive roll, pitch braking, and yaw burst. The shared viewport should visibly switch between full-assembly and frame-only geometry, stress and displacement fields, and replay motion.
  4. Open Review → Branches, Evidence, and History. The result must remain an interactive estimate, and Use this frame must remain disabled because no automatic promotion occurs.
  5. For a local source verification, run pnpm test:run and pnpm build after pnpm install with the pinned Node and pnpm versions.

The September 3 production recording exercises this complete deployed flow: WebMCP-controlled assembly selection and validation, a 0.35-material topology target, four named structural-estimate cases, signed displacement deformation at an explicitly labelled 1,000x scale, and a final human-only review boundary.

Public Demo

https://webmcp-structural-evolution.vercel.app

Public Repo

https://github.com/SebastianBoehler/webmcp-structural-evolution

The public repository includes the Apache-2.0 LICENSE and THIRD_PARTY_NOTICES.md.

Demo Video

Video URL: https://youtu.be/pjcF61w_8M4

Screenshot Shot List

  1. Agent-synchronized reference assembly: the human and agent share the visible FPV-drone workbench after context inspection.
  2. Agent-generated estimate: the deployed dashboard shows the reviewable Balanced frame estimate produced by generate_topology_candidate.
  3. Human review with disabled promotion: the human reviews the estimate while Use this frame remains disabled at the estimate stage.

Readiness

The public demo returns successfully without credentials, the public repository is available with an Apache-2.0 license, and the finished narrated demo is under three minutes at https://youtu.be/pjcF61w_8M4. Submission 1167572 was submitted to The WebMCP Challenge and verified live on September 3, 2026 at https://devpost.com/software/structural-evolution.

Known Limitations

  • Dashboard topology output is an interactive sparse-SIMP lattice estimate, not certified FEA or manufacturing approval.
  • It does not provide free-form B-rep/STEP synthesis, nonlinear/contact/kinematic solving, CFD, thermal/fatigue validation, experimental validation, or flight approval.
  • WebMCP depends on a client exposing document.modelContext; WebGPU-dependent paths require a compatible browser/device.
  • There is no backend, remote compute, collaboration service, accounts, or durable project storage in this release.

Built With

Share this project:

Updates

Submission history