modelanalyzer-mcp
# modelanalyzer-mcp
An [MCP](https://modelcontextprotocol.io/) server that lets your coding assistant recommend, compare and price AI models, and check when they retire, using the free open data from [AI Model Analyzer](https://modelanalyzer.ai).
It runs locally over stdio and reads one public file, `https://modelanalyzer.ai/data/summary.json`, which the site rebuilds every night. There is no API key, no account and no server to host. Answers link back to the matching page on the site.
## Install
Add it to your assistant's MCP configuration:
```json
{
"mcpServers": {
"modelanalyzer": { "command": "npx", "args": ["-y", "modelanalyzer-mcp"] }
}
}
```
For Claude Code:
```bash
claude mcp add modelanalyzer -- npx -y modelanalyzer-mcp
```
Requires Node.js 20 or later.
## Tools
| Tool | What it does |
| --- | --- |
| `list_scenarios` | The use-case scenarios that `recommend_models` ranks for |
| `recommend_models` | The site's ranking for a scenario (quality blended with cost at a typical workload), optionally under a monthly budget or open weights only |
| `search_models` | Find models by name, id or maker, with Analyzer Index, price, cost per task, speed and context |
| `compare_models` | Side-by-side table of 2 to 6 models |
| `estimate_cost` | Cost per request, day and month at list price for a token volume |
| `model_retirements` | Announced API shutdown dates with each maker's recommended replacement |
Example prompts: "Which model should I use for a customer support bot under $100 a month?", "Compare Claude Opus 5.5 and GPT-6.1 Sol", "Is any model I use being retired soon?"
## Data and credit
Numbers are exactly what modelanalyzer.ai shows, generated nightly. Cost per task and speed come from [Artificial Analysis](https://artificialanalysis.ai/); Arena scores from [LMArena](https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset) (CC-BY 4.0). See the [methodology](https://modelanalyzer.ai/about/) and [open data](https://modelanalyzer.ai/open-data/) pages.
The summary is cached in memory for six hours. Set `MODELANALYZER_BASE_URL` to read from another host (for example a local build).
## Development
```bash
npm install
npm test
```
Bug reports and ideas are welcome in [issues](https://github.com/yeaske/modelanalyzer-mcp/issues). The rankings and numbers themselves come from the site; see its [methodology](https://modelanalyzer.ai/about/).
## Releasing
Bump the version in `package.json`, `server.json` (twice) and `src/server.js`; `npm test` checks they agree. Then:
```bash
npm publish
mcp-publisher login github
mcp-publisher publish
```
The second step lists the new version in the [MCP Registry](https://registry.modelcontextprotocol.io/) as `io.github.yeaske/modelanalyzer-mcp`.
## License
MIT, © Digital Designs Now.
TDQS
Scored across 6 tools
Each tool maps to a distinct operation: enumeration of scenarios, ranking, name-based lookup, multi-model comparison, cost math, and retirement dates. search_models vs compare_models are the closest pair, but the descriptions clearly separate 'find by name' from 'side-by-side table of 2-6 models.'
Five of six tools follow a clean verb_noun snake_case pattern (list_scenarios, recommend_models, search_models, compare_models, estimate_cost). model_retirements breaks the pattern with a noun-only form, but it remains readable and unambiguous.
Six tools is a tight, well-scoped surface for a model-analysis domain, with no filler or redundant endpoints. Each tool covers a distinct analytical need an agent would actually ask for.
Discovery, ranking, comparison, cost estimation and lifecycle coverage are all present, forming a coherent read-only pipeline. Minor gaps: no single-model detail view distinct from search, and no way to list all tracked models or filter by capability/context constraints beyond budget and open weights.