The ty\e API

Overview

The ty\e API turns a text description or a photo of a part into real, editable CAD — an analytic B-rep STEP file (AUTOMOTIVE_DESIGN / AP214) you can open in any CAD/CAM package. It is a small REST API built on three endpoints:

Simplest — one call: /cad/from-image or /cad/from-text (the server runs vision/chat + build for you, and returns the STEP file). For more control — to inspect or edit the recipe before building — chain the building blocks: /chat/vision or /chat (→ recipe) then /cad/step (→ STEP).

• Request body: application/json
• /chat & /chat/vision response: text/event-stream (SSE)
• /cad/step response: application/step (a downloadable STEP file)
• Authentication: Authorization: Bearer gx_...
• Available on plans: Pro · Max

Authentication

Every request must include a Bearer token in the Authorization header. Keys are 56-character hex strings prefixed with gx_.

Authorization: Bearer gx_a4f8e2b1c9d7f3e5a2c8b1d4e6f9c3a7b061e2345678901abcdef23

Get a key

1. Sign in at tylellm.com/app
2. Upgrade to Pro or Max
3. Settings → API → Show or Copy

Rotate a key

If a key leaks, click Regenerate in the API panel. The old key is invalidated immediately. There's a hard rate-limit of 3 regenerations per 5 minutes per user.

Models

Available model IDs

Pass one of these as the model field. Vendor model IDs change often — GET /models returns the live list, and the byok:provider:served passthrough targets any served name not listed here.

Engineer roadmap

Where each ty\e tier is heading. Timelines are best-effort and may shift as model training completes.

POST /cad/from-image

One call: a part photo becomes a STEP file. The server runs the vision model to get a recipe, then builds it — you send the image, you get geometry back. The simplest way from a photo to CAD.

Request body

Response

200 with Content-Type: application/step — the STEP file (analytic B-rep, AP214) + headers X-Tye-Faces, X-Tye-Bodies, X-Tye-Build-Ms. Errors: 400 (image download failed), 422 (no recipe / unbuildable), 502/503 (vision or build unavailable). One request runs vision + build, so expect ~30–60s — use a long client timeout. Rate-limited to about 30 requests/min per IP (429 when exceeded); if the builder is saturated you may get 503 builder busy — retry shortly.

Example request

{
  "image":    "https://example.com/part.jpg",
  "model":    "gm_31_pro",
  "byok_key": "YOUR_PROVIDER_KEY",
  "name":     "part"
}

POST /cad/from-text

One call: a text description becomes a STEP file. The server runs the chat model to get a recipe, then builds it.

Request body

Response

Same as /cad/from-image — 200 application/step + X-Tye-* headers, ~30–60s.

Example request

{
  "prompt":   "Design a mounting bracket, 100x60x6 mm, with four M8 holes",
  "model":    "gm_31_pro",
  "byok_key": "YOUR_PROVIDER_KEY"
}

POST /cad/step

Build a recipe into a real analytic-B-rep STEP file — the endpoint that returns geometry. Feed it a <tye-cmd> recipe (from /chat or /chat/vision) or a raw feature list.

Request body

Response

200 with Content-Type: application/step — the STEP file bytes (analytic B-rep, AP214). Headers: X-Tye-Faces, X-Tye-Bodies, X-Tye-Build-Ms. Errors: 400 (no recipe/features), 422 (recipe failed to build), 503 (build service busy).

Example request

{
  "features": [
    { "op": "spurGear", "params": { "module": 2, "teeth": 24, "width": 12, "bore": 10 } }
  ],
  "name": "gear"
}

POST /chat/vision

Reverse-engineer a part photo into a parametric <tye-cmd> recipe, streamed as SSE. Requires a signed-in (Pro/Max) key and a vision model.

Request body

The reply streams tokens (same {token, done} format as /chat) and ends with a <tye-cmd>{...}</tye-cmd> recipe. Pass that recipe to /cad/step for the STEP file.

POST /chat

Turn a text prompt into a recipe — or hold a normal engineering conversation. With a build prompt (e.g. “Design a bracket…”) and a strong model, the streamed reply contains a <tye-cmd> recipe you can pass to /cad/step.

Request body

Message roles

• user — your prompt
• assistant — previous AI reply (for multi-turn context)
• system — optional system override (rarely needed; Custom Instructions handle this server-side)

Example request

{
  "messages": [
    { "role": "user", "content": "Design a mounting bracket, 100x60x6 mm, with four M8 holes" }
  ],
  "stream": true,
  "temperature": 0.3
}

Streaming format

When stream: true, /chat and /chat/vision return text/event-stream. Each line is data: <json>\n\n where the JSON is {"token": "...", "done": false}. Read line by line, strip the data: prefix, and concatenate every token.

data: {"token": "Here", "done": false}
data: {"token": " is", "done": false}
data: {"token": " <tye-cmd>...", "done": false}
…
data: {"token": "", "done": true}

The stream ends with {"token": "", "done": true}. Set stream: false to receive the full reply as one JSON object instead.

Errors

Errors come back as standard HTTP status codes with a JSON body.

{ "detail": "Invalid API key" }

Rate limits

Limits are per-account, not per-key. Regenerating a key does not reset the daily counter.

When you hit a limit the server returns 429 with a Retry-After header indicating seconds to wait. Implement exponential backoff in production clients.

Code samples