Every published calculator is an API — and an MCP tool

Language models guess at arithmetic. A published MetaModel project answers with computed results — schema discovery plus server-side evaluation by the same dependency-tracked formula engine the interactive page runs. No SDK, no keys for published calculators, no hallucinated numbers.

MCP — plug calculators into any AI assistant

One hosted endpoint speaks the Model Context Protocol: https://www.metamodel.app/api/mcp — add it to claude.ai as a custom connector (Settings → Connectors), or point any MCP client at it. Three tools:

metamodel_list_projects

Browse the published calculators available to compute against — name, description, and token per project.

metamodel_get_schema

Discover a calculator's contract: every input with its type, defaults and valid ranges; every output; per-model. 3D presentation models are flagged so agents know what's engineering and what's display.

metamodel_compute

Send inputs, get every output computed server-side by the same dependency-tracked formula engine the published page runs — formulas, lookups, rules and all.

Prefer to run it yourself? The server is open source: github.com/metamodel-app/mcp-server.

REST — two calls, no auth for published calculators

These examples run against a live project — the hydraulic cylinder configurator, a real engineer-to-order sizing chain (buckling, stress, standard-part selection, costing). Paste them in a terminal right now.

1 — Discover the contract (inputs, types, ranges, outputs):

curl https://www.metamodel.app/api/compute/d5279c53-ceb3-4b26-b2ac-71caf2e3c3ae

2 — Compute:

curl -X POST https://www.metamodel.app/api/compute/d5279c53-ceb3-4b26-b2ac-71caf2e3c3ae \
  -H "Content-Type: application/json" \
  -d '{"inputs": {"bore": 63, "stroke": 500, "pressure": 160, "mountStyle": "Clevis"}}'

Response (excerpt — every value computed server-side):

{
  "models": {
    "cylinder": {
      "outputs": {
        "forceKn": 49.9,      // push force from p·A
        "minRodDia": 34.8,    // Euler buckling ∨ direct stress
        "rodDia": 36,         // next standard rod (threshold lookup)
        "barrelOd": 75,       // hoop-stress wall
        "weightKg": 16.7,     // real part masses
        ...
      }
    },
    "quote": {
      "outputs": { "totalCost": 634, "sellPrice": 813, ... }
    }
  }
}

An AI reviewed this model over MCP — and found a real engineering gap

An AI assistant exploring the cylinder configurator through these exact tools noticed the rod-sizing indicator only guarded one direction. Digging in revealed the real issue: at short strokes, buckling stopped governing and the model could select a rod past its compressive yield. The fix — a direct-stress floor alongside the Euler term — shipped the same day, verified by hand-computed truths. That's the loop these APIs exist for: AI reads the schema, computes real cases, and reasons about engineering it can trust — because the numbers come from the engine, not the language model.

Good to know

A publish token is a public read-and-compute capability — anyone with the token can fetch the schema and compute. Unpublishing revokes it.
Schema and results flag 3D presentation models (presentation: true) so agents can separate engineering outputs from display geometry.
Compute is metered per plan with generous free daily quotas — fair use, not a paywall surprise.
There's an llms.txt for AI assistants reading this site.
Authoring APIs (create and edit projects programmatically) are on the roadmap — today's surface is read + compute.

Ship your own compute endpoint

Build a calculator — formulas, lookups, rules, even 3D — publish it, and it's instantly an interactive page, a REST API, and an MCP tool. One model, three surfaces.

Start building See the live demos