Six productsInstall one or all. One sign-in (Microsoft, Google or GitHub) works for all six.
SAFE MCP is a free MCP server that lets AI assistants such as Claude, Codex and Cursor read and work with the slab or foundation model open in CSI SAFE on your PC — slabs, mats, footings, soil supports, post-tensioning and design results — with a preview before every risky change.
Run the analysis, then show the maximum soil pressure and the largest deflection under SERV1.
Illustrative session. The assistant works on the model already open in your SAFE.
Tools for everyday slab and foundation work, plus the whole SAFE API reflected from the SAFE you use, with typed, risk-checked calls. 239 capabilities are catalogued, each with its state and whether it was verified in a real SAFE.
Query slabs, drops, ribbed and waffle slabs, mats, footings, beams and supports, filtered by story, group or selection.
Create the analysis model (mesh) without running, and set mesh options through SAFE's tables.
Define soil subgrade springs, compression-only too, and audit supports: unsupported models, foundations without soil, unused springs.
Tendon properties, and a tendon audit with profile, strands, jacking and losses. Tendons are read-only through SAFE's API.
Run the analysis; it is reported finished only when every case is. Long runs work in the background.
Strip-based slab design and its strip results, governing strips, and punching-shear checks sorted by ratio.
Soil pressures under mats and footings, the largest slab deflections, and SAFE's slab detailing with top and bottom bars.
Read, bulk-edit, import and export SAFE's tables, for anything no tool covers.
A SAFE plug-in cannot host a server, so SAFE MCP is a program of its own: your AI assistant starts it, and it attaches to the SAFE already open and calls the SAFE API — the one installed with your SAFE, loaded when it runs.
With several SAFE windows open, the assistant asks which one you mean — it never guesses. When SAFE is waiting for the CSI licence sign-in, SAFE MCP says so: sign in to SAFE in its window, and the tools answer again.
SAFE MCP has no public release yet. This page shows the download and the install steps by itself once it is out.
safe-mcp.exe --doctor checks SAFE, its API and your AI clientssafe-mcp.exe --setup all connects every installed AI clientSetup connects every installed client below, and installs the SAFE skills for Claude; safe-mcp.exe --setup all does it again at any time. The workflow skills are also served as MCP prompts, so clients without skill support still get them.
Setup connects every installed client for you; nothing to type.
safe-mcp --cloud on.https://aecomcps.com/mcp/safeSign in with the same Microsoft, Google or GitHub account as on your PC, start a new chat with SAFE MCP switched on, and keep SAFE open. What the AI reads from the model passes through the team gateway to the AI company; the gateway never stores or logs it.
A slab or foundation model needs correctness and a trail more than speed. Every change is checked before and after it is made.
Changes accept a dry run, so the assistant can show what would change before it does.
Deleting, unlocking results, a new, opened or imported model, applying tables or overwriting files always shows a preview and needs your yes; the confirmation token is single-use and bound to that model.
Each change names the model it is meant for (expectedModel) and is refused if another model is open.
Switch it on in the settings and every tool that changes the model, the application or files is hidden and refused.
Every change reads the affected items back and reports whether it is verified; a failed check is reported, not hidden.
Loads, soil values, post-tensioning forces and design codes come from you: the assistant asks. Codes, combinations and assumptions change only when you ask.
The saved model is backed up before high-impact steps, and an operation log records every change and preview.
Your model stays on the PC unless you switch on Cloud AI access for claude.ai or ChatGPT, and the gateway never stores what passes. When you sign in, only privacy-filtered AI learning is shared with your team. Logs and backups stay in %APPDATA%\MCPs\SafeMcp.
Analysis, design, punching-shear, soil-pressure and detailing results are software output for the engineer's review, not proof that the structure is adequate or code-compliant.
SAFE MCP works through the SAFE API of CSI SAFE 2020 to 2026 (SAFE 20 to 26), using the SAFE installed on your PC. It was verified live in SAFE 23.3.1: 97 capabilities, write self-test 101 of 104 steps.
Tendons are read-only there: no adding, jacking or deleting tendons through the API (SAFE's tendon tables or SAFE itself do it). Geometry editing is limited to moving objects and changing connectivity: no replicate, divide or merge.
Not possible through the SAFE API, so not offered: direct ETABS .EDB import, DXF/DWG and IFC exchange, drawings and schedules, SAFE's own report writer, the tendon layout generator, display settings, and cancelling a running analysis.
SAFE MCP is a free Model Context Protocol (MCP) server for CSI SAFE, so AI assistants such as Claude, Codex, Cursor and VS Code Copilot can read and work with the slab or foundation model open in SAFE on your PC through the SAFE API.
CSI SAFE 2020 to 2026 (SAFE 20 to 26) on Windows. It was verified live in SAFE 23.3.1.
Query slabs, mats and footings; create the analysis mesh; define and audit soil and spring supports; audit post-tensioning tendons and their losses; run the analysis; and read strip and slab design, punching shear, soil pressure, detailing and deflection results and SAFE's database tables.
Not through SAFE's API: in SAFE 23 tendons are read-only there, so SAFE MCP reads and audits them (profile, strands, jacking and losses). Tendons are added, jacked or deleted through SAFE's tendon tables or in SAFE itself.
No. SAFE MCP never invents loads, soil values, post-tensioning forces or design codes: the assistant asks you. Analysis, design and detailing results are SAFE's software output for the engineer's review.
No. Your model, its names and its values stay on your PC. When you sign in, SAFE MCP shares only what the AI learns about using the tools, with names and numbers removed. Only if you switch on Cloud AI access, what claude.ai or ChatGPT reads from the model passes through the gateway, which never stores it.
Claude Code, Claude Desktop, Codex, Cursor, VS Code (Copilot), Windsurf, Gemini CLI and other MCP clients on your PC, connected by setup, with the SAFE skills for Claude. ChatGPT and claude.ai in the browser connect through the optional Cloud AI access (safe-mcp --cloud on), off unless you switch it on.
Yes. SAFE MCP is free. You need a licensed CSI SAFE 2020 to 2026 on Windows 10 or 11 (64-bit).
SAFE MCP has no public release yet. The download appears on this page by itself once it is out.
Guides: All guides
Revit, Navisworks and 3ds Max are trademarks of Autodesk, Inc. ETABS and SAFE are trademarks of Computers and Structures, Inc. Photoshop is a trademark of Adobe Inc. STAAD.Pro and OpenSTAAD are trademarks of Bentley Systems, Incorporated. AECO MCPs is an independent product and is not affiliated with Autodesk, Computers and Structures, Inc., Adobe or Bentley Systems.