A quiet orb that reads your canvas and only speaks when it's caught a real, verified failure. A one-click fix on your live graph — never a rebuild. And underneath, the build engine that places native, self-healing Grasshopper definitions across any model.
A standing, read-only loop reads every solve that already happened — never triggering its own — so it always knows what's solving and what silently isn't.
A peripheral dot reads your canvas continuously — data shapes, types, empty outputs — and never interrupts the solve. You never explain your file.
The silent killer: a graph that turns green and produces nothing. MANTIS sees the empty output the moment a rewire breaks it.
It reads the real result of your live solve, so "this is empty and here's why" is a machine-checked fact — not a guess. Only verified catches earn a word.
One click reconnects the wire on your actual graph — one undo, neighbours untouched — then reads back: "fixed, now produces 96 panels."
15–45s screen capture of a real streaming build — prompt → reasoning → native nodes appear → click a plan step to jump. Drops in here.
One segmented composer, four intents — Build is the default. The partner that catches your silent breaks is also a fluent builder when you want one.
Describe a design in plain English. MANTIS plans it, streams the reasoning, then places native Grasshopper components on your canvas — wired, grouped by stage, and labelled. Click any plan step to frame its group.
A design assistant across both Rhino and Grasshopper. It creates and edits geometry, runs Rhino commands, and builds definitions step by step — recording undo points and asking before anything destructive.
A conversational answer about Grasshopper or your current canvas. MANTIS reads what's there and explains or advises — no building. The fastest way to learn the parametric move you're missing.
Iterate builds on top of the existing graph — add panelization, floor thickness, mullions, nothing dropped. Multi-Solution generates 2–3 alternative approaches so you can compare before committing.
Hand MANTIS a reference photo, a sketch, or a drawing with a dimension on it. A vision model reads the form and the numbers — a "6000" callout becomes a 6000 mm reference — then builds a parametric definition scaled to match. Not a traced mesh: a real, editable graph.
# facade_elevation.jpg → read grid 12 × 8 bays, mullion @ 1500mm scale dim callout "6000" → 6000mm build surface → bay grid → mullion sweep verify 96 panels · 0 warnings · editable
A folded plan emits before the JSON — you see how MANTIS stages the graph and why, in one streamed call. Educational, not decorative.
After building, it reads back live component values, detects warnings and type mismatches, and repairs them — then reports a verdict.
Each plan step maps to a labelled group on the canvas. Click a step and MANTIS frames that exact cluster of components.
It remembers what worked in your file — a lessons flywheel that makes later builds in the same project sharper.
Runs end with a chip: green "verified" or amber "issues remain." You always know the health of what it built.
Every action records an undo point. Destructive operations ask first. Nothing happens to your model you can't reverse.
Five providers, latest generation each — or run fully offline.
Sonnet & Opus over HTTP, or the Claude CLI for local-key streaming with vision.
GPT-4.1, GPT-4.1 Mini, o3-mini — JSON-mode reliable, vision-capable.
2.5 Flash (free) and 2.5 Pro, 1M-token context for big catalogs.
One key, many models — a curated shortlist of the strongest builders.
Fully local, no key, no network. Your prompts never leave the machine.
Switch providers per-task from the panel. No MANTIS account, no telemetry gate.