meshcheck

Getting started

Get a key, run your first validation, and read the report card.

Get a key

Everything below needs an API key. Getting one is self-serve and takes a minute:

  1. Pick a plan on the pricing page. Free ($0, 300 credits/month) needs no card; paid plans (Indie, Studio, Scale) run on Polar, our merchant of record.
  2. Sign up at /signup with your email. We send a single-use sign-in link — no password. (A paid plan carries you straight to checkout after you sign in.)
  3. Mint a key from your dashboard. The raw key is shown once, right then — we store only a hash and can’t recover it. If you lose it, mint a new one and revoke the old. Keys are never sent by email.

The site’s drag-and-drop demo needs none of this: it runs on a separate, server-held demo key you never see, rate-limited per IP.

Authentication

Every request carries your key in the X-Api-Key header:

X-Api-Key: mc_live_...

No key returns 401 MISSING_API_KEY; a bad key returns 401 INVALID_API_KEY. Keys map to an account with a credit balance and a per-plan rate limit.

Your first validation

Send a model as multipart/form-data with a file field. Pick a budget profile (web, mobile, pc, or hero); it sets the thresholds the checks measure against.

curl -X POST https://api.meshcheck.dev/v1/validate \
  -H "X-Api-Key: mc_live_..." \
  -F file=@barrel.glb \
  -F profile=web

Files up to the platform’s multipart cap (~4.5 MB) post directly. For larger models, send a JSON body with a public url (the server fetches it, with an SSRF guard), or use the presigned upload flow: POST /v1/uploads returns a single-use upload_url and a blob_id, you PUT the bytes, then validate by blob_id. Models up to 20 MB validate synchronously; above that, the call returns a job envelope you poll.

Reading the report card

The response is a single JSON document. Three fields carry the verdict:

  • verdict — the rollup: pass, warn, or fail. Any error-severity check that fails makes the whole report fail; any warn and no failures makes it warn; otherwise pass. There is no numeric score.
  • summary — counts of errors, warnings, info, and skipped results.
  • checks[] — one CheckResult per check, in a fixed ID order, each with structured measured and threshold objects and a human-readable message derived from them.
{
  "verdict": "fail",
  "summary": { "errors": 1, "warnings": 2, "info": 6, "skipped": 0 },
  "checks": [
    {
      "id": "PERF-001",
      "name": "Triangle count",
      "status": "fail",
      "measured": { "triangles": 48231 },
      "threshold": { "max_tris": 25000 },
      "message": "48,231 triangles exceeds profile budget of 25,000"
    }
  ]
}

An agent acts on status and the structured measured/threshold values — never by parsing the message prose. Because the core is deterministic, the same file and profile always produce the same report (timestamps and signed URLs aside), so you can cache and diff reports across regenerations.

Determinism and the one exception

Every /validate check is a pure measurement. The only non-deterministic surface is /inspect, which asks a vision model a semantic question (“does this look like a barrel?”). It is a separate endpoint, priced separately, and every answer is labelled deterministic: false with the model version. Vision opinions never appear in a /validate report.

Next steps

  • The Checks reference — every check ID, what it measures, and its default severity.
  • The API reference — every endpoint, option, error code, and credit cost.
  • The Report schema — the field-by-field contract your code builds against.
  • The MCP guide — drive meshcheck from Claude Code or any MCP client.