StructureClaw

AI-assisted structural engineering workspace for AEC workflows.

Quick Start · Try It · Why · Engines · Architecture · Docs · 中文

Demo

StructureClaw-demo.mp4

Quick Start

No Node.js Yet

Use the bootstrap installer when the machine does not already have Node.js 20+ and npm:

curl -fsSL https://raw.githubusercontent.com/structureclaw/structureclaw/master/scripts/install.sh | bash

Windows PowerShell:

irm https://raw.githubusercontent.com/structureclaw/structureclaw/master/scripts/install.ps1 | iex

The installer lets interactive users confirm or edit StructureClaw Home before it prints the final install plan. Press Enter to keep ~/.structureclaw, or type another path to change the workspace/runtime directory. It reuses existing Node.js 20+ installations and prepares a user-local Node.js 24 runtime when needed, installs @structureclaw/structureclaw@latest, adds the user-local npm bin directory to PATH, and runs sclaw doctor. On Windows, the bootstrap Node runtime is installed under %LOCALAPPDATA%\Programs\nodejs; on Linux, it is installed under ${XDG_DATA_HOME:-~/.local/share}/nodejs.

Install from npm

If Node.js 20+ and npm are already installed:

npm install -g @structureclaw/structureclaw
sclaw doctor
sclaw start

Open the local workspace printed by sclaw start. sclaw doctor creates the runtime workspace, checks LLM settings, prepares SQLite, and installs the Python analysis environment when needed.

Run from source

git clone https://github.com/structureclaw/structureclaw.git
cd structureclaw
./sclaw doctor
./sclaw start
./sclaw status

Windows PowerShell from source:

node .\sclaw doctor
node .\sclaw start
node .\sclaw status

China mirror entrypoint:

sclaw_cn doctor
sclaw_cn start

Try It In 60 Seconds

After the workspace opens, try:

Model a two-story steel frame, 6 m by 4 m bay, 3.6 m story height, Q355 steel columns and beams, dead load 5 kN/m2 and live load 2 kN/m2. Build the model, run static analysis, check GB50017, and generate a short report.

Expected flow:

draft model -> validate -> run analysis -> code-check -> report

Use OpenSees for a fully open local run. Select PKPM or YJK only when the corresponding commercial software and authorization are available on the machine.

Why StructureClaw

StructureClaw turns a natural-language structural description into a traceable engineering workflow:

describe -> draft model -> validate -> analyze -> code-check -> report

What makes it useful:

• Chat-first modeling: describe a frame, truss, portal frame, or generic structure and let the agent build a computable model.
• Real analysis handoff: run OpenSees, PKPM SATWE, or YJK through the same backend-hosted analysis contract.
• Traceable engineering artifacts: keep model drafts, validation results, tool calls, analysis outputs, checks, and reports visible.
• Local-first runtime: installed mode stores data in the user runtime directory instead of the package directory.
• Extensible skills and tools: combine built-in skills with user-local skills/tools under the runtime workspace.

Engine Support

Engine	Skill	Best for	Requirements
OpenSees	opensees-*	Open, repeatable static/dynamic/seismic/nonlinear analysis	Python analysis environment prepared by sclaw doctor
PKPM SATWE	pkpm-static	Commercial-engine static checks and SATWE comparison	Local PKPM installation, JWSCYCLE.exe, valid license
YJK 8.0	yjk-static	YDB conversion, YJK static calculation, structured result extraction	Local YJK 8.0 installation, bundled Python 3.10, valid authorization

Commercial engines are explicit selections and require local software installation. StructureClaw does not bundle PKPM or YJK.

Architecture

flowchart LR
  User[Engineer in chat UI] --> Web[Next.js workspace]
  Web --> API[Fastify backend]
  API --> Agent[LangGraph agent runtime]
  Agent --> Skills[Skill layer]
  Agent --> Tools[Tool layer]
  Tools --> Analysis[Python analysis runtime]
  Analysis --> OpenSees[OpenSees]
  Analysis --> PKPM[PKPM SATWE]
  Analysis --> YJK[YJK 8.0]
  Tools --> Reports[Reports and artifacts]

Loading

Main directories:

• frontend/: Next.js 14 application
• backend/: Fastify API, agent/chat flows, Prisma integration, and analysis execution host
• scripts/: startup helpers and the sclaw / sclaw_cn CLI implementation
• tests/: regression runner (node tests/runner.mjs ...), install smoke, and CI-covered frontend checks (type-check, Vitest, lint) after native smoke
• docs/: user handbook and protocol references

Runtime Modes

Mode	Command	Data directory	Process model
bootstrap installer	scripts/install.sh / scripts/install.ps1	user runtime directory plus user-local Node/npm prefix	installs Node if missing, then installs the npm package
npm install	sclaw start	user runtime directory, defaulting to ~/.structureclaw/	backend serves the exported frontend in one process
source checkout	./sclaw start	user runtime directory, defaulting to ~/.structureclaw/	backend and frontend run as development processes

Recommended local flow:

./sclaw doctor
./sclaw start
./sclaw status

China mirror flow (same subcommands, mirror defaults enabled):

./sclaw_cn doctor
./sclaw_cn start
./sclaw_cn status

Bootstrap installer mirror flow:

curl -fsSL https://raw.githubusercontent.com/structureclaw/structureclaw/master/scripts/install.sh | bash -s -- --cn

& ([scriptblock]::Create((irm https://raw.githubusercontent.com/structureclaw/structureclaw/master/scripts/install.ps1))) -Cn

Notes:

• scripts/install.sh and scripts/install.ps1 are intended for first-time users. They reuse existing Node.js 20+ installations and install Node.js 24 into a generic user-level Node.js directory if node / npm are missing or too old, then run npm install -g @structureclaw/structureclaw@latest.
• The bootstrap installers use a user-local npm prefix by default (~/.structureclaw/npm-global), avoiding root/admin global npm writes.
• SQLite is now the default local database. ./sclaw start uses ~/.structureclaw/data/structureclaw.start.db, and ./sclaw doctor uses ~/.structureclaw/data/structureclaw.doctor.db so preflight checks do not touch the active local runtime database.
• ./sclaw doctor no longer requires a preinstalled system Python 3.12. It will ensure uv and prepare a virtual environment with Python 3.12 automatically when needed. On Windows, this automatic setup currently requires winget; if winget is unavailable, install uv manually before running ./sclaw doctor.
• If your old local .env still points DATABASE_URL at a local PostgreSQL instance, ./sclaw doctor and ./sclaw start will auto-migrate that data into SQLite, rewrite .env to the SQLite default, and keep the original PostgreSQL URL in POSTGRES_SOURCE_DATABASE_URL.
• That first auto-migration also creates a local backup file like .env.pre-sqlite-migration.<timestamp>.bak.
• sclaw_cn defaults to China mirror settings when unset: PIP_INDEX_URL=https://pypi.tuna.tsinghua.edu.cn/simple, NPM_CONFIG_REGISTRY=https://registry.npmmirror.com.
• You can override mirror values in .env or shell environment (PIP_INDEX_URL, NPM_CONFIG_REGISTRY, APT_MIRROR).

Useful follow-up commands for source checkouts:

./sclaw logs
./sclaw stop
# Requires a source checkout:
node tests/runner.mjs backend-regression
node tests/runner.mjs analysis-regression

Use the built-in CLI batch convert command to transform structure model JSON files and write a summary report:

./sclaw convert-batch --input-dir tmp/input --output-dir tmp/output --report tmp/report.json --target-format compact-1

Windows PowerShell:

node .\sclaw doctor
node .\sclaw start
node .\sclaw status
node .\sclaw logs all --follow
node .\sclaw stop

Configuration

StructureClaw 1.0 uses settings.json as the user-facing configuration file. sclaw doctor creates it, and the General Settings panel writes the same settings through the backend admin API.

Configuration resolution:

1. Runtime settings.json
2. Built-in defaults

Selected environment variables still participate as runtime fallbacks or directory controls. PORT, FRONTEND_PORT, and NODE_ENV are read when the corresponding setting is absent, and SCLAW_DATA_DIR changes the runtime directory used to locate settings.json and data files.

Important settings.json fields and sections:

• server: ports, host, request body limit
• llm: OpenAI-compatible base URL, model, API key, timeout, retries
• database.url: SQLite connection URL
• logging: application log level, LLM logging, log rotation
• analysis: Python runtime path, timeout, engine manifest path
• storage: reports directory and upload size
• agent: workspace root, checkpoints, shell-tool policy
• pkpm: SATWE/JWSCYCLE path and work directory
• yjk: install root, executable, bundled Python, work directory, version, timeout, headless mode

By default, settings.json lives in ~/.structureclaw/. The SCLAW_DATA_DIR environment variable can override that runtime directory for tests or controlled deployments.

API Entrypoints

Backend:

• POST /api/v1/agent/run
• POST /api/v1/chat/message
• POST /api/v1/chat/stream

Backend-hosted analysis:

• POST /validate
• POST /convert
• POST /analyze
• POST /code-check
• GET /engines

Engineering Principles

• Skills are enhancement layers, not the only execution path.
• Unmatched selected skills fall back to generic no-skill modeling.
• User-visible content must support both English and Chinese.
• Keep module boundaries explicit across frontend/backend/analysis skills.

Documentation

• English handbook: docs/handbook.md
• Chinese handbook: docs/handbook_CN.md
• Documentation hub: docs/README.md
• English reference: docs/reference.md
• Chinese reference: docs/reference_CN.md
• Chinese overview: README_CN.md
• Contribution guide: CONTRIBUTING.md
• Roadmap: ROADMAP.md
• Security policy: SECURITY.md

Contributing

Please read CONTRIBUTING.md and the organization-level Code of Conduct before opening a PR.

License

MIT. See LICENSE.