AI Model Watch

ai-model-watch

List the models each AI provider's API serves, and see what's new since the last time you looked. One command, no dependencies, Node 18+. It also runs as an MCP server, so Claude or Cursor can check for you.

New models often show up in a provider's /models list before any announcement. ai-model-watch saves each provider's list, and on the next run prints only the model IDs that weren't there before. Run it by hand, from cron, or in CI with --exit-code.

$ ai-model-watch -p openrouter

2026-10-17T09:26:01.231Z OpenRouter: 2 new

• deepseek/deepseek-v4.1-flash-1017 DeepSeek: DeepSeek V4.1 Flash 1017 (listed as created 2026-10-17)
• reflection/beam-1.0-501b-a23b Reflection: Beam 1.0 501B A23B (listed as created 2026-10-16)

The first run records a baseline (OpenRouter: 470 models recorded as the baseline), so you only hear about changes after that.

Providers

Listing models is free on every one of these APIs. Nothing is generated and no tokens are spent. A provider whose key isn't set is skipped. Keys are sent only to that provider, never across a redirect, and never printed.

Install and run

brew install markksantos/tap/ai-model-watch   # macOS / Linux (Homebrew)

or, with Node 18+

npx github:markksantos/ai-model-watch

or

git clone https://github.com/markksantos/ai-model-watch && node ai-model-watch/bin/ai-model-watch.js

ai-model-watch [options] -p, --providers openrouter,openai,anthropic,gemini,mistral,xai,deepseek (default: all with a key) -s, --state where the last-seen lists live (default: ~/.cache/ai-model-watch/state.json) -w, --watch keep running and check every N seconds (minimum 60) --removed also print models that disappeared --json JSON output --exit-code exit 1 when anything new appears (for cron and CI) -q, --quiet print only new models --mcp run as an MCP server on stdio

MCP server (Claude, Cursor, VS Code)

--mcp runs it as a Model Context Protocol server on stdio, so an assistant can ask what's new itself. Two tools:

• check_new_models: the model IDs each provider started listing since the last check (same state file as the CLI).
• list_models: what one provider lists right now, with an optional substring filter.

Claude Code:

claude mcp add ai-model-watch -- npx -y github:markksantos/ai-model-watch --mcp

Or as a Claude Code plugin (this repo is its own marketplace):

/plugin install ai-model-watch --marketplace markksantos/ai-model-watch

On Claude Code versions older than 2.1.275, run /plugin marketplace add markksantos/ai-model-watch, then /plugin install ai-model-watch@ai-model-watch. The plugin passes any provider keys set in your environment through to the server, and adds an ai-model-watch skill so Claude reaches for the tools when you ask whether a model is out yet.

The skill on its own (Claude Code, Cursor, Codex and other agents), from claude-plugins.dev:

npx skills-installer install @markksantos/ai-model-watch/ai-model-watch

Claude Desktop, one click: download ai-model-watch-1.2.0.mcpb and open it. Desktop asks for the optional keys in its settings.

Claude Desktop, Cursor and other clients (mcpServers in their JSON config):

{

"mcpServers": { "ai-model-watch": { "command": "npx", "args": ["-y", "github:markksantos/ai-model-watch", "--mcp"] } } }

OpenRouter works without a key. To include another provider, add its key from the Providers table to an env object in the same entry.

Cron

*/15 * * * * OPENAI_API_KEY=... ANTHROPIC_API_KEY=... ai-model-watch -q >> ~/model-drops.log

GitHub Actions

examples/github-action.yml runs the check on a schedule, caches the state file between runs, and fails the job (so GitHub emails you) when a new model appears.

Library

npm install github:markksantos/ai-model-watch, then:

const { check, fetchProvider } = require('ai-model-watch');

const report = await check({ providers: ['openrouter', 'anthropic'] }); for (const p of report.providers) { for (const m of p.added || []) console.log(p.label, m.id); }

fetchProvider(id, { env, fetch }) returns { status: 'ok' | 'skipped' | 'error', models: [{ id, name, created }] }. check() diffs against the state file and saves the new lists. A provider that errors keeps its last good list. Errors include HTTP and network failures, a redirect, a body that isn't that provider's list shape, a malformed model ID, and pagination that doesn't finish. So a bad response never empties the saved list or makes the next run report the whole catalog as new.

What a list can't tell you

A model ID in a /models list means the provider lists it. It doesn't prove the model answers requests yet, and some providers list models days before they are generally available. created is whatever date the provider reports, not when the model appeared.

The hosted version, [ArtificialWatch](https://artificialwatch.com), polls these APIs plus the public catalogs every 60 seconds and confirms a detection on a second sweep before anything fires. Then it alerts you by email and Chrome push (free), or by text or phone call. Use it if you'd rather not keep a cron job running.

License

MIT