Model Context Protocol

MCP server for AI model data

Your coding agent's knowledge of models stops at its training cutoff, so it writes model IDs that are outdated or retired. Connect this server and it can look up what is current instead: model IDs, VRAM per quantization, API prices, benchmarks and daily download trends for 1,423 open-weights and API-only models.

Endpoint

Streamable HTTP

https://aimodelscomparison.com/mcp

Free, no sign-up and no API key. Read-only: the tools only look things up.
Rate limit: 60 tool calls per minute per IP.
Data refreshes nightly. Every answer carries data_as_of, attribution and a source_url per model for citation.

Connect

Claude (claude.ai and desktop): Settings → Connectors → Add custom connector, name it “AI Models Comparison”, paste the endpoint URL, leave authentication empty. ChatGPT: in developer mode, add a connector with the same URL. For the rest, copy a snippet:

Claude Code

claude mcp add --transport http aimodelscomparison https://aimodelscomparison.com/mcp

Cursor · ~/.cursor/mcp.json

{
    "mcpServers": {
        "aimodelscomparison": {
            "url": "https://aimodelscomparison.com/mcp"
        }
    }
}

VS Code · .vscode/mcp.json

{
    "servers": {
        "aimodelscomparison": {
            "type": "http",
            "url": "https://aimodelscomparison.com/mcp"
        }
    }
}

MCP Inspector (test it)

npx @modelcontextprotocol/inspector
# Transport: Streamable HTTP · URL: https://aimodelscomparison.com/mcp

Tools

Each tool, with a question it answers.

resolve_model_id — Resolve a current model ID

“What is the current model ID for Claude Sonnet? Is gpt-4o still available?”

What it returns

Use BEFORE writing code that calls a model, or to check whether a model name/ID is still current. Accepts a slug, Hugging Face id, OpenRouter id (":free"/":batch" and date suffixes OK) or free text. Returns status (active/expiring/deprecated/removed), callable ids (OpenRouter + variants, Hugging Face), release date, knowledge cutoff and the newest model in the same family. Ambiguous input returns did_you_mean instead of guessing. Example: {"query": "claude sonnet"}

whats_new — Recent model releases, removals and price changes

“Which models were released or removed this month? Did any API prices drop?”

What it returns

Use when asked about recent AI model releases, models removed from APIs, or API price changes — anything after your training cutoff. Returns new_models (by release date), removed_models (delisted from OpenRouter) and price_changes (≥10% move, old/new USD per 1M tokens, % delta), newest first. Filters: days (1–90), type, deployment_type, task (pipeline_tag). Example: {"days": 30, "type": "new", "deployment_type": "api_only"}

find_models — Find models by requirements

“Cheapest coding model under $1 per 1M input tokens with tool calling?”

What it returns

Use to choose a model by requirements: task, budget, context, features, licence, or what fits your GPU. Covers open-weights and API-only models. Returns compact rows: ids, params, context, prices (USD per 1M tokens), Artificial Analysis intelligence index, 30-day downloads, plus a VRAM fit verdict when fits_vram_gb/hardware is set. sort=value is intelligence per blended dollar (3:1 input:output). Example: {"task": "coding", "max_price_input_usd_per_m": 1, "supports": ["tools"], "sort": "value"}

get_model — Full model profile

“Everything about Llama 3.1 8B Instruct — licence, VRAM, price, trend.”

What it returns

Use for everything we know about ONE model: specs, licence and commercial-use verdict, status/freshness, API pricing (incl. :free/:batch variants and price tiers in plain words), benchmarks, VRAM per quantization with a verdict per hardware preset, daily download trend (Δ7d/Δ30d, ≤30-point sparkline), price history and family (base model, top derivatives). Trim with include. Example: {"model": "llama 3.1 8b instruct", "include": ["vram", "trend"]}

compare_models — Compare 2–4 models side by side

“Compare Claude Sonnet with DeepSeek V4.1 Flash.”

What it returns

Use to compare 2–4 models (open-weights and API-only can be mixed). Returns an aligned table — the same fields for every model, missing values spelled out with a reason ("n/a — API-only") — plus the list of fields that differ, the canonical compare page URL when one exists for a pair, and its cached auto-generated verdict if there is one. Example: {"models": ["claude sonnet", "deepseek-ai/DeepSeek-V4.1-Flash"]}

can_i_run — Can I run this model locally?

“Can I run Qwen3 8B on an RTX 3060? Which quantization?”

What it returns

Use when someone asks whether a model runs on their GPU/Mac. Give a hardware preset or vram_gb (+ unified_memory for Apple silicon), optionally context_tokens. Returns per quantization: file size, estimated VRAM and a verdict (runs | tight | no), plus the recommended quant. Without hardware, returns verdicts for every preset. API-only models get a clear "cannot self-host" answer with the cheapest API option. Estimates, not benchmarks. Example: {"model": "qwen3 8b", "hardware": "rtx3060-12"}

estimate_cost — Estimate monthly API cost

“What would 50M input + 5M output tokens a month cost on GPT-6 Luna?”

What it returns

Use to estimate monthly spend for a token volume. Returns cost via OpenRouter's main price, the cheapest provider endpoint, and :batch/:free variants when they exist (cached-input share billed at the cache-read price). For open-weights models also returns the minimum hardware preset that runs it — no self-host dollar estimate. Example: {"model": "claude sonnet", "input_tokens_per_month": 50000000, "output_tokens_per_month": 5000000, "cached_input_ratio": 0.6}

trending — Fastest-growing open models

“Which open text-generation models are gaining downloads fastest?”

What it returns

Use when asked which open-weights models are gaining adoption. Ranks by growth of Hugging Face 30-day downloads over 7d or 30d, from our own daily snapshots — the same ranking as aimodelscomparison.com/trending (one row per model family). Returns downloads_30d and the absolute and % change. Falls back to download volume while history is shorter than the window (growth_ranking=false). Example: {"period": "7d", "task": "text-generation"}

Also available: resources aimc://hardware-presets and aimc://methodology, and the prompt choose_model, which walks the agent through find → compare → resolve.

What makes it different

Model freshness: release dates, retirements, delistings, and the newest model in each family, so an agent stops writing IDs from its training data.
Can I run it locally: estimated VRAM per quantization, with a runs/tight/no verdict for common GPUs and Macs.
Self-host vs API: API cost for your token volume next to the hardware needed to run the weights yourself.
History: daily download snapshots since 28 Jul 2026 and a daily API price history. Hugging Face and OpenRouter only show the current value.

Data sources & terms

Hugging Face Hub API — model metadata, licences, file inventories, download and like counters for open-weights models. Our daily snapshots of those counters are our own dataset.
OpenRouter API — API-only models, prices per 1M tokens, :free/:batch variants, provider endpoints, release dates, knowledge cutoffs and retirement dates.
Artificial Analysis (intelligence, coding and agentic indices) and Design Arena ratings — third-party benchmarks surfaced via OpenRouter. We display them; we do not run them.

Derived figures (VRAM estimates, trends, value scores, licence classes) are ours and documented on this page. When you reuse data from the site, the JSON API or the MCP server, credit “Data: Hugging Face API, OpenRouter API, Artificial Analysis and Design Arena via OpenRouter. Compiled by aimodelscomparison.com” and link the model's page. Provider data remains subject to its source's terms; every value is a best-effort copy refreshed daily, not a guarantee of current list prices.

Prefer plain HTTP? The same data is in the /api/v1 JSON API and summarised in llms.txt. Methods: methodology.