Steltic

Free, open-source AI agent that designs steel buildings to AISC 360 and AISC 341 (with AISC 358 where appropriate) and ASCE 7-22 loads, on your own machine. Paste a design brief and get an OpenSees model, a calculation report written for senior review, and an interactive 3D viewer. Bring your own LLM API key (your API base URL and key are held in memory only, never written to disk).

• Easiest start: Steltic Hub installs this and the other modules in one window.
• See the output first: 36 worked buildings with reports and 3D models at stelticai.com.
• Videos: initial design run · optimize and finalize.
• Indian Standards edition: steltic_india.

browser ──▶ FastAPI app (localhost) ──▶ your LLM (key in app memory, never stored)
                 │  parses tool calls
                 ▼
           sandbox executor (run_python only) — Docker when available
                 │
           steel engine + OpenSees ─▶ report.html + viewer_3d.html

Not for construction. Every output is produced by your own AI model, may be incomplete or incorrect, and must be independently checked and sealed by a licensed professional engineer before any use for design, construction, or permitting. See DISCLAIMER.md.

Install & run

Most people should start with Steltic Hub, which clones and installs this module for you. To run this module on its own, install it from GitHub. It needs Python 3.10–3.12 (openseespy, the analysis engine, ships compiled binaries for those versions only); 3.12 is recommended.

Windows (PowerShell)

git clone https://github.com/Steltic/steltic
cd steltic
py -3.12 -m venv .venv
.venv\Scripts\Activate.ps1
pip install -e .
steltic                      # starts on http://127.0.0.1:8000 and opens your browser

macOS / Linux

git clone https://github.com/Steltic/steltic
cd steltic
python3.12 -m venv .venv
source .venv/bin/activate
pip install -e .
steltic                      # starts on http://127.0.0.1:8000 and opens your browser

The install brings the full engine (openseespy, numpy, scipy, matplotlib, about 400 MB), so designs run without Docker. To run with a .env file of settings, use ./run_local.sh instead of steltic.

Update: in the repo folder, git pull, then pip install -e . again (needed only when pyproject.toml changed, harmless otherwise).

First run: open Settings, enter your provider's API base URL (any OpenAI-compatible endpoint — OpenRouter, vLLM, Together, Fireworks, OpenAI, Anthropic…), your API key, and a model id. Then paste a brief (stories, bays & spacings, loads, seismic, system) — or pick one of the 36 built-in example briefs — and click Design building.

Offline smoke test: set Model to MOCK — it drives the whole pipeline (sandbox, engine, OpenSees, report) with no LLM.

Windows note: if starting fails with "Application Control policy has blocked this file" or "os error 4551", Windows Smart App Control is blocking an unsigned Python. Use the signed python.org build and recreate the environment:

winget install Python.Python.3.12       # then open a new terminal in the repo folder
Remove-Item -Recurse -Force .venv
py -3.12 -m venv .venv
.venv\Scripts\Activate.ps1
pip install -e .

Sandbox

The agent's Python only ever executes in a sandbox. EXECUTOR (env / .env) picks the mode:

value	isolation	notes
auto	Docker if available, else subprocess	default
docker	container per run: no network, read-only fs, non-root, cpu/mem caps	build once: ./sandbox_image/build.sh
subprocess	none (child process with rlimits)	needs pip install openseespy numpy scipy matplotlib; fine for local single-user use

Steltic binds to 127.0.0.1 and has no authentication — don't expose the port. Designs are saved under your OS user-data dir (e.g. %LOCALAPPDATA%\Steltic, ~/.local/share/Steltic); override with DATA_DIR.

Engineering-standards, OpenSees & design-examples RAG (highly recommended)

Steltic was validated with RAG vector databases of AISC 360, 341 and 358 (2022), the AISC Steel Construction Manual Design Examples (V16.0, 168 worked examples), an examples database of validated OpenSees models, and two databases of the OpenSees documentation. The specifications and design examples ground the LLM in the current procedures; the OpenSees databases help it build the model — and when OpenSees throws errors, resolve them. Steltic's performance has been validated with these databases and will likely reduce without them.

For copyright reasons the vector databases of AISC 360, 341 and 358 are not open-sourced in this project. The rest are freely downloadable from this repo's v0.1.1 release as portable .jsonl.gz dumps (pre-embedded; see rag_v2/README.md for loading them into your own Qdrant): steel_design_examples (168 original worked Q&A covering the AISC Design Examples scope), opensees_buildings_3d (40 validated 3D building models), opensees_building_templates, and the two OpenSees documentation sets.

Two options:

1. None (default — not recommended). Leave RAG_API_URL empty — the agent relies on its own cited AISC knowledge. Good models do respectably; grounded runs are better.
2. Bring your own server. Run any server exposing the small API described in rag_v2/README.md: load the downloadable OpenSees databases, and build the standards collections from your own licensed copies with the rag_v2/ chunking + ingestion starter kit. Then put the connection in the environment (or a .env next to where you launch): RAG_API_URL=http://your-server:8080/query and, if your server enforces one, RAG_API_TOKEN=....

Feedback from the Nonlinear module (SNL)

The Nonlinear module can hand a re-design back through Continue once its Chapter 16 / pushover / DDM run is complete (see steltic_nonlinear/docs/README_feedback.md). Three things in this repository exist for that:

• cfg['drift_relief_16_1_2'] — the ASCE 7-22 §16.1.2 relief block SNL writes when a Chapter 16 analysis lets the §12.12.1 drift limits be relaxed (Risk Category I–III only). steel_engine/preflight.py and consistency.py accept a relaxed drift_limit only with a complete block, never for Risk Category IV, and never above the Chapter 16 limit; Chapter 8 of the report states the relief and that the design is provisional until Chapter 16 is re-run.
• POST /api/restore/<building>?archive=1 — moves an existing job aside as <building>__<timestamp> before the restore (what a fresh run already does), so a promoted candidate never silently overwrites a design of record; viewer_3d.html is now part of the restore whitelist.
• design/design_of_record.json — written by SNL when the user promotes a verified candidate; the report's Chapter 1 basis table shows it. The agent contract (contract/AGENT_START.md, Feedback from the Nonlinear module) says how the three loop briefs must be applied (capacity_design.SCWB.by_story, capacity_design.panel_zone.by_joint).

tests/test_drift_relief.py covers the rules (python -m pytest tests -q).

Build-only mode: a building the user designed

For a steel building the user designed, HR Steel builds the analysis package for the Nonlinear module without designing anything (steltic/build_only/). Nothing is resized and nothing is invented:

1. Intake -- the Steltic building template (GET /api/build_only/template, one tab per group: project and basis, grid, levels, columns, beams, braces, loads, details, wind, your own results) or an ETABS .e2k text model becomes one building_description.json. What the .e2k states is taken as data (stories, grids, frame objects per story, sections, local-axis angles, releases, base restraints, diaphragms, design preferences); what has to be interpreted (the dead and live load per level from area loads + slab + framing self-weight, the system per direction when the design preferences name one frame type for a building that also has braces) is a proposal with its breakdown.
2. Missing items (missing_items.json) -- key items have no default: HR stops and asks (a section that is not an AISC shape is one question per section name, the AISC metric designations W360X64 / HSS254X254X9.5 resolve by themselves); default items show the proposed value and its reason and are used only once accepted; refused items are outside version 1 (plate shear walls, STMF, partially restrained connections, sloped or plan-diagonal beams, members off the levels, several diaphragm types, per-member grades, springs) and are never simplified.
3. Translate -- translate.py writes cfg.py with the user's members as literal tables and a custom_build that replays them one for one (add_column / add_beam / Truss, the diaphragms, attach_diaphragm_mass). What it derives (grid lines through off-grid columns, column and beam splits where members frame in, the floor plate, Table 12.12-1 drift limit, Cu, Ct/x, Kd, GCpi) is listed in provenance.json with its rule, beside the source of every other value (file / answer / accepted default).
4. Build -- the normal pipeline (model_opensees.py, model_static.py, report.html, design/*). calc_package.json marks every member not checked (the engine has no coded capacities; a user design is not checked by Steltic HR) and carries the user's RBS cuts, Lb and doublers for the Nonlinear module.

The agent never redesigns such a job: it gets contract.BUILD_ONLY_RULES, its write_file refuses the generated files, and the completion gate is the build-only gate (runner.gate: cfg.py is the generated one, the model is built as declared, nothing is open). HTTP: POST /api/build_only/intake | answers | build, GET /api/build_only/status/<building>. In the hub the User-designed building module drives the whole flow. steltic/build_only/extract.py re-enters an HR job as a description (the acceptance test: 11 of the 18 HR test buildings re-enter with identical periods, base shear and seismic weight -- and identical pushover / NLRHA models in the Nonlinear module; the other 7 are refused for version-1 reasons: plate shear walls, STMF, R = 3 not specifically detailed, plan-diagonal chamfer beams, inclined transfer columns, crane brackets off the levels). tests/test_build_only.py.

A tool call the model never finished

A provider behind OpenRouter validates the whole conversation on every request. If a turn is cut off in the middle of a tool call's arguments ({ and nothing more), a provider such as Together refuses the next request with 400 Invalid JSON in tool call arguments — a 400 the retry net does not retry, so the run ended, and the cut-off turn saved in conversation.json broke every Continue after it. The agent now keeps every tool call in the history parseable ({} stands in for arguments that did not parse), never runs such a call, answers it with an error that asks the model to repeat it with complete arguments, and heals a conversation saved by an older version when it is resumed. tests/test_tool_call_healing.py.

Repo map

steltic/ FastAPI app + agent loop + sandbox executors · steel_engine/ OpenSees modelling, design pipeline, consistency checks, report + 3D viewer · contract/ the agent's working contract and references · frontend/ vanilla JS UI · sandbox_image/ Docker sandbox image · test_buildings/ 50+ example briefs (steel + CFS) with assessment rubrics · rag_v2/, rag_update/ build-your-own-RAG starter kit.

License

MIT — see LICENSE, NOTICE (third-party data notes) and DISCLAIMER.md (engineering disclaimer; also shown in-app at /terms).