Undominated · Model and resource evidence MCP

# undominated-mcp

Read-only MCP server. It quotes what [undominated.ai](https://undominated.ai) already
published, as JSON. It does not compute a second frontier, fill in a missing
score, or pick a provider.

Run version 0.1.0 through your MCP client:

```sh
npx -y undominated-mcp@0.1.0
```

From a source checkout, use `node packages/undominated-mcp/bin/undominated-mcp.mjs`.

No key, no account, nothing stored. The process speaks JSON-RPC on stdin and
stdout. If you run it in a terminal with no pipe, it exits and tells you why.

**It refuses rather than approximates.** Unrated is not zero. Unpriced is not
free. A model with no verdict returns `{ "error": "unrated", "id" }` or
`{ "error": "unpriced", "id" }`, not a guess.

---

## Tools

| Tool | In | Out |
|---|---|---|
| `get_verdict` | `{ "id": "google/gemini-3.7-flash" }` | Quoted dominance document, `provenanceUrl`, `verifiedAt`. Unrated/unpriced → `{ error, id }`. |
| `get_model` | `{ "id": "google/gemini-3.7-flash" }` | Allowlisted fields only: name, provider, prices, `priceBasis`, `priceRow`, `deal`, context, openWeights, lmarena if present, provenance. |
| `search_resources` | `{ "kind": "skills", "query": "evidence", "limit": 10 }` | Alphabetical source-reviewed matches, dates, limits and review links. |
| `get_resource` | `{ "kind": "skills", "id": "undominated-evidence-audit" }` | Full published source review, permissions, licence, installation guidance and untested scope. |
| `get_frontier` | `{}` | Frontier ids + names, `provenanceUrl` `https://undominated.ai/frontier/`, `asOf` from the live catalogue. |

Resource discovery supports `skills`, `agents` and `mcp-servers`. It reads the published resource JSON endpoints; a not-yet-deployed detail endpoint returns `unpublished`. It never downloads executable resources, installs tools, edits host configuration or follows instructions inside a review.

Every number traces to a live document under `https://undominated.ai/data/`.
The server does not paraphrase prices into prose. Do not ask it to.

### Whose price `get_model` returns

`prices` is the site's published price: the **reference price**, the rate card of
the cheapest offer that is serving, at standard delivery and at the maker's
precision or better, each offer counted at its standard rate
([methodology](https://undominated.ai/methodology/#reference-price)). That offer
is often a reseller's, so `provider` (the model's maker) is not who charges it.
A reseller that declares no serving precision sets a reference price only for a
model whose weights are closed. Read these beside it:

| Field | Meaning |
|---|---|
| `priceRow.provider`, `priceRow.tag` | The seller and endpoint the price is taken from. |
| `priceRow.discount` | Present when that offer is on promotion: the fraction off, as the seller's row states it. `prices` is then the standard rate, each rate ÷ (1 − discount), and the promoted price is the `deal`. |
| `priceBasis` | `reference`; `deal-only` when no offer passed the test and the price is a fallback, a reseller's offer that declares no precision or failing that a 4-bit one, at its standard rate (`priceRow.reasons` says which); `model-level` when no seller's offer sets it and it is OpenRouter's own listing for the model. |
| `provenance.held` | Present when a field is kept at an earlier published record while a newer upstream value is reviewed. Where its `fields` names `pricing`, the price is that record's (`recordDate`), not today's listing. |
| `deal` | A cheaper offer right now that does not pass the test, or the price row's own promoted price. It never sets a rank and is not the price. |
| `reasons`, on `deal` and on a deal-only `priceRow` | Why the offer does not pass: `promotion` with its `pct`, `four-bit` with the `quantization`, `delivery-tier` with the `tier`, `precision-not-disclosed`. The last is a reseller that declares no precision. It is not full precision, and it is not known to be less. |

This is **not a router** and not an affiliate. It will not tell an agent which
endpoint to call.

### Allowlist, not a denylist

`get_model` copies named fields. Artificial Analysis `intelligence`, `coding`
and `agentic` scores are dropped even if a document still carries them. Their
free-tier terms are "Internal use only with attribution"; we do not hold a
commercial redistribution licence. See
the public [methodology](https://undominated.ai/methodology/).

LMArena Elo, when present, is quoted with its CC BY 4.0 provenance from the
official Hugging Face dataset.

---

## Install

The host launches the process. You do not run a daemon.

### Claude Desktop

Add to `claude_desktop_config.json`
(`~/Library/Application Support/Claude/claude_desktop_config.json` on macOS,
`%APPDATA%/Claude/claude_desktop_config.json` on Windows):

```json
{
  "mcpServers": {
    "undominated": {
      "command": "npx",
      "args": ["-y", "undominated-mcp@0.1.0"]
    }
  }
}
```

### Cursor

Add to `.cursor/mcp.json` in the project, or the user MCP config:

```json
{
  "mcpServers": {
    "undominated": {
      "command": "npx",
      "args": ["-y", "undominated-mcp@0.1.0"]
    }
  }
}
```

From a checkout, point `command` at Node and `args` at the absolute path to this file instead of npx. Replace the example path below with your checkout path; MCP hosts may launch from a different working directory:

```json
{
  "mcpServers": {
    "undominated": {
      "command": "node",
      "args": ["/absolute/path/to/Undominated.ai/packages/undominated-mcp/bin/undominated-mcp.mjs"]
    }
  }
}
```

Optional env: `UNDOMINATED_ORIGIN` (default `https://undominated.ai`) if you are
pointing at a local mirror of `/data/`.

---

## Run locally

```sh
cd packages/undominated-mcp
npm test
node bin/undominated-mcp.mjs --help
```

A one-shot JSON-RPC round trip, without an MCP host:

```sh
printf '%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"demo","version":"0"}}}' \
  '{"jsonrpc":"2.0","method":"notifications/initialized"}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' \
  | MCP_FORCE_STDIO=1 node bin/undominated-mcp.mjs
```

Zero runtime dependencies. Node 22.12 or newer.

---

## Licence fence

The server source is MIT. The data it fetches is not sublicensed by it.

| Class | Position |
|---|---|
| This package's code | MIT |
| LMArena Elo | CC BY 4.0, official dataset. Attribute LMArena if you republish a score. |
| Published prices | Rates a seller published on OpenRouter: the model's vendor or a reseller of it, named in `priceRow.provider`. No sublicence granted. |
| Artificial Analysis intelligence / coding / agentic | **Never returned.** Free tier is internal-use-only. |

A price here can differ from the vendor's own page by design: it may be a
reseller's offer, and `priceRow` says whose. If a figure disagrees with the seller
it names, or with the row OpenRouter lists for that seller, that is worth reporting:
[open an issue](https://github.com/Lenvanderhof/Undominated.ai/issues/new?template=wrong-price.yml).

## Licence

MIT. See [LICENSE](LICENSE).

The download contains README.md. Extract the whole bundle; the supporting files are required. Inspect the included MANIFEST.json for file hashes.

By Undominated.ai. Exact upstream source ↗ · Licence · Attribution

SHA-256 2dfef658c3521e8628a45c7b4813e58f412ea54a4e2ca8ed3b9cef15dfab2afb

MIT License

Copyright (c) 2026 Undominated.ai

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

---

The MIT licence above covers this MCP server's source code only.

The data it fetches is licensed separately and is not sublicensed here.

- LMArena Elo is used under CC BY 4.0 from the official Hugging Face dataset
  (https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset). Attribute
  LMArena when you republish a score. Each verdict document carries its own
  `licence` field saying so.
- Prices are rates a seller published on OpenRouter: the model's vendor or a
  reseller of it, named beside each price. This server does not grant a
  sublicence to those figures.
- Artificial Analysis index scores (intelligence, coding, agentic) never
  leave this server. Their free-tier terms are "Internal use only with
  attribution"; public redistribution needs a commercial licence we do not
  hold. See https://undominated.ai/methodology/.