Free beta — evaluation use

SGT Intelligence MCP Server

An MCP server for AI agents building in Blender.

Connect your agent

Streamable HTTP · Bearer API key

https://mcp.sgtfleet.com/mcp

llms.txt · Full agent guide · MCP card · HTTP API · Catalogue JSON

Start with 14 days of Premium access, then keep the Free plan forever. No card. Watermarked previews and a draft evaluation licence, not reviewed by counsel.

Start the free 14-day trial Quickstart

Choose your starting point

Free to start. Room to grow.

Forever

Free

Chair, table and bicycle. 3 build attempts per UTC day. GLB export and watermarked previews.

Start with the trial

On signup · No card

14-day trial

All 972 permitted families. 25 build attempts per UTC day. GLB and FBX. Watermarked previews, then the Free plan forever.

Start the free 14-day trial

Coming soon

Premium

Higher limits, no watermark, and a commercial licence once reviewed by counsel.

Tell us what you'd build

Free and trial: evaluation only under draft terms, not reviewed by counsel. One trial per email address. Failed attempts count. No prices announced.

$ benchmark --show-evidence

Benchmarks

Historical paired experiments. Measured gains, with the costs and failures kept in view.

About 6× fewer defects

gpt-6.1-sol (frontier model), TF1 · n = 24 paired briefs · p = 0.03125. Mean defects 0.17 library vs 1.04 raw. Dimension adherence 100% in both arms. Method and source

4.5× fewer defects with a small model

gpt-6-luna (smaller, cheaper model), TF2 STANDARD · n = 24. Mean defects 6.4 library vs 28.7 raw; descriptive, no significant paired defect test established. Method and source

About 1/9 the cost per finished complex asset

TF2 COMPLEX · 10 briefs per arm. gpt-6-luna + library $0.024 vs raw gpt-6.1-sol $0.22. List-price estimate, not an invoice; 7 vs 9 finished. Method and source

Blind COMPLEX gpt-6-luna library vs raw gpt-6-luna: mean +0.5, one-sided lower bound +0.1 (n = 10); significant on this secondary set.

Limitations. Raw gpt-6.1-sol was preferred in blind STANDARD visual judging over gpt-6-luna + library (mean -0.625). No quality-parity claim. TF1 library arms used 1.73× (gpt-6.1-sol) and 2.07× (gpt-6-luna) as many weighted tokens per finished asset as raw arms. Neither study met its full pre-registered win criteria; some measurements were missing. Retained protocol failures limit interpretation.

Methods, full numbers and limitations · benchmarks.json

Find a starting point

Search the library.

Ranked search across names, descriptions, tags, presets and parameter names. Try all 972 permitted families during your 14-day trial; chair, table and bicycle stay available on Free.

Search families

Loading catalogue…

Explore the library

Showcase.

A few of the 972 object families in the library. Each sheet shows the default and three variants from the same function.

Loading examples…

Premium interest

Tell us what you would build with Premium. Leave an optional build idea and consent to updates, or contact support@sgtfleet.com. No unsolicited emails or third-party trackers.

Free beta — evaluation use

Request a trial key.

Automatic 14-day trial on signup, no card: all 972 permitted families, 25 build attempts per UTC day, GLB and FBX. Then the Free plan forever: chair, table and bicycle, 3 attempts per UTC day and GLB. One trial per email address. Previews stay watermarked; evaluation licence only, draft and not reviewed by counsel.