Ollama Library Scraper

Returns: one row per Ollama model with name, description, capabilities, sizes, pulls, tagCount, updated and sourceUrl. Input: {"maxItems":20} Best for: local-LLM catalogues, tracking model popularity over time, finding models by capability or parameter size.

{"name": "llama3.2", "capabilities": ["tools"], "sizes": ["1b", "3b"], "pulls": "85.4M", "tagCount": 63, "updated": "2 years ago", "sourceUrl": "https://ollama.com/library"}

What does Ollama Library Scraper do?

This Ollama scraper provides an Ollama API workflow through Apify and collects structured data about the AI models listed in the public Ollama model library. You can list the whole library, whose model count changes over time, or run a keyword search such as "llama", "qwen" or "coder". Every model becomes one clean record that you can export as JSON, CSV, Excel or HTML, or read through the Apify API.

The Actor reads only public pages. It needs no login, no account and no API key. It is designed for people who want to track which open models exist, compare their sizes and capabilities, watch pull counts grow, or feed a model catalogue into a spreadsheet, a database or an AI agent.

Typical uses:

Build a catalogue of local LLMs with their parameter sizes and capability tags.

Monitor which models are popular by tracking pull counts over time.

Find every model that supports a given capability, for example vision or tool calling.

Detect new or recently updated models by checking the update date.

Give an AI agent a live list of models it can pull and run locally.

This tool is an independent scraper. It is not affiliated with, endorsed by or sponsored by Ollama.

What data can you get?

Each record describes one model from the library. The table lists every field in the output.

Field	Type	Description
id	string	Stable ID, the same as the model name
name	string	Model name as listed in the library, for example llama3.1
url	string	Library page of the model
sourceUrl	string	The page the record was scraped from (the library listing or the search page)
description	string	Short description shown in the listing
capabilities	array	Capability tags such as tools, vision, thinking, embedding, cloud
sizes	array	Parameter-size tags such as 8b, 70b, 405b
pulls	string	Pull count as the site shows it, for example 120M
pullsNumeric	number	The pull count converted to a number (approximate, because the site abbreviates it)
tagCount	number	Number of tags (variants) of the model
updated	string	Relative last-update text, for example "1 year ago"
tags	array	Only when includeTags is true: per-tag digest, download size, context window, input types and update age.
updatedAt	string	Exact last-update time in ISO 8601, UTC

Search-page records include pulls and pullsNumeric from the download badge. They can have null tagCount, updated and updatedAt when those values are not present in the listing markup. Download badges are excluded from capabilities. Library records usually include them. A null value means unavailable, not zero; agents should check it before comparisons.

Some models have no capability tags, so capabilities is an empty list for them. A few models have no size tags either, in which case sizes is empty. Embedding models and cloud-hosted models are tagged accordingly in capabilities.

How to use

1. Open the Actor in Apify Console and go to the Input tab.
2. Optionally type a keyword into Search query. Leave it empty to read the whole library.
3. Optionally pick one or more Required capabilities. Only models that carry all the chosen tags are kept.
4. Set Max models to limit the output. The Actor never outputs more records than this number.
5. Click Start. A run over the full library finishes after fetching the selected public pages.
6. Open the Output tab to view the table, or download the data in your preferred format.

You can also start the Actor from code, from a schedule, from an integration such as Make or Zapier, or through the MCP server described below.

How searching works

With an empty search query the Actor reads the full library listing. With a search query it reads the site's own search results page, so the ranking is the site's ranking. The site search page supplies its top-ranked matches. When more results are needed, or capability filters are supplied, the Actor also reads the library and adds models whose name, description or capability text contains every query word. The maxItems limit applies after filtering and sorting. Leave the query empty to browse the library without keyword matching.

If the search finds nothing, for example because the keyword is nonsense, the run finishes successfully with zero items and logs a warning and sets a final status suggesting a broader search, fewer capabilities or a lower minPulls. It does not output placeholder records.

Select a model watchlist

Use modelNames when you already know the models you want to monitor. It accepts a mixture of plain library identifiers and full public Ollama library URLs. For example, {"modelNames":["llama3.2","https://ollama.com/library/qwen3:8b"],"maxItems":20} selects the two parent model records. A tag suffix identifies a variant, but the dataset still contains one row per model, not one row per tag. Turn on includeTags to inspect the variants within the model's tags array.

Exact selection avoids the ambiguity of keyword search. Searching for llama may find several related names; selecting llama3.2 keeps only the matching library name. Repeated names or URLs do not create duplicate rows. A missing name is reported in the log, while other names in the watchlist can still produce records. The selection is checked against the public library listing, so community uploads outside that listing are not included.

The searchQuery, capability, parameter-size and minimum-pull filters still apply to a watchlist. When modelNames is supplied, site search is skipped and every searchQuery word must match the selected library model. If a selected model does not satisfy those filters, it will be omitted. The global maxItems limit applies across the entire watchlist. This makes the input useful for scheduled comparisons: keep the same list between runs and compare each stable id, approximate pullsNumeric and updatedAt in your own storage. The Actor does not retain a historical series or calculate changes between runs.

Capability filter

The capability filter is applied after the page is read. It uses the capability tags shown for each model. When you select several capabilities, a model must have all of them. For example, selecting vision and tools keeps only models that are tagged with both.

To shortlist smaller variants, use {"parameterSizes":["1b","3b"],"sortBy":"pulls","maxItems":20}. The filter keeps models with at least one matching parameter-size badge. It combines with keyword, exact-name, capability and minimum-pull filters; those other filters must also match. The complete sizes array remains in the output, so a model can include larger variants too. This selects parent models rather than individual tags. Use includeTags when you need the listed variants and their download sizes. Parameter count alone does not establish memory requirements or guarantee that a model will run on your hardware. Missing size badges cannot match this filter.

Input

Field	Type	Description
parameterSizes	array	Match any parameter-size badge, e.g. ["1b", "3b"]. Case insensitive; default [] applies no size filter.
modelNames	array	Exact model names or library URLs. Tagged URLs select the parent model. Empty by default.
searchQuery	string	Keyword to search for, for example llama. Empty means the whole library. Search may be supplemented with matching library models.
capabilities	array	Allowed values: tools, vision, thinking, embedding, cloud. A model must carry all selected tags.
maxItems	integer	Maximum number of models to output, between 1 and 1000. Default 50.
sortBy	string	default (site order), pulls (most pulled first) or updated (most recently updated first).
minPulls	integer	Keep only models with at least this many pulls. 0 means no filter.
includeTags	boolean	Default false. When true, each model also gets a tags array with tag, digest, size, contextWindow, inputTypes and updated, read from the model's tags page. This adds one request per model.
proxyConfiguration	object	Apify Proxy settings. Default direct requests; block recovery attempts Apify Proxy when available.

Example input:

{
  "searchQuery": "llama",
  "capabilities": ["tools"],
  "maxItems": 20
}

Start with {"maxItems":20} to get library rows without tag requests. All filters combine; maxItems is one shared limit after filtering and sorting. An empty exact selection tells you to check modelNames spelling or URL, clear searchQuery, remove capability or size filters, or lower minPulls. An unknown capability has no matching models and returns zero results. Malformed URLs and invalid limits fail input validation with a clear message.

Output

The Actor writes one dataset item per model. Here is a real item from a run:

{
  "id": "llama3.2",
  "name": "llama3.2",
  "url": "https://ollama.com/library/llama3.2",
  "sourceUrl": "https://ollama.com/library",
  "description": "Meta's Llama 3.2 goes small with 1B and 3B models.",
  "capabilities": [
    "tools"
  ],
  "sizes": [
    "1b",
    "3b"
  ],
  "pulls": "85.4M",
  "pullsNumeric": 85400000,
  "tagCount": 63,
  "updated": "2 years ago",
  "updatedAt": "2024-09-25T21:09:00.000Z"
}

Records are unique by id. The default dataset view shows the most useful columns: name, description, capabilities, sizes, pulls, tag count, update text and the model URL.

How much does it cost?

The Actor is charged per result: you pay for each model record that is saved to the dataset. Check the Pricing tab of the Actor page for the current rates. The standard library listing uses one page request. Enabling tag details adds a page request per selected model. You can keep the total spend predictable by setting Max models.

Use with AI agents (MCP)

You can call this Actor from any MCP-compatible AI assistant through the Apify MCP server. Add the following configuration to your client:

{
  "mcpServers": {
    "apify": {
      "url": "https://mcp.apify.com/?tools=axiomworks/ollama-library-scraper"
    }
  }
}

Example prompts:

"List the Ollama models that support vision and show their sizes."

"Which Ollama models matching 'qwen' have the most pulls?"

"Get the 20 most recently updated Ollama models using sortBy updated and show their updatedAt values."

"Find Ollama embedding models and give me their descriptions."

The agent sees the input fields of the Actor, so it knows it can pass a search query, capability filters and a maximum number of items.

Integrations and API

You can run the Actor through the Apify API and read the results as soon as the run finishes.

Python

from apify_client import ApifyClient

client = ApifyClient("YOUR_APIFY_TOKEN")
run = client.actor("axiomworks/ollama-library-scraper").call(
    run_input={"searchQuery": "llama", "maxItems": 20}
)
for item in client.dataset(run["defaultDatasetId"]).iterate_items():
    print(item["name"], item["pulls"], item["sizes"])

JavaScript

import { ApifyClient } from 'apify-client';

const client = new ApifyClient({ token: 'YOUR_APIFY_TOKEN' });
const run = await client.actor('axiomworks/ollama-library-scraper').call({
    capabilities: ['vision'],
    maxItems: 25,
});
const { items } = await client.dataset(run.defaultDatasetId).listItems();
items.forEach((m) => process.stdout.write(`${m.name}\n`));

cURL

curl -X POST "https://api.apify.com/v2/acts/axiomworks~ollama-library-scraper/run-sync-get-dataset-items" \
  -H "Authorization: Bearer $APIFY_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"searchQuery": "llama", "maxItems": 20}'

You can also connect the dataset to Google Sheets, Slack, webhooks, Make, Zapier and other integrations available on the Apify platform, or schedule the Actor to run daily and compare pull counts between runs.

FAQ

Can I paste a model URL into searchQuery? Yes. Public library URLs are normalized to their model name; a tagged URL such as /library/llama3.2:3b searches for its parent model. For exact matching or several targets, use modelNames. Search URLs read their q parameter, including encoded text.

Does one unavailable model fail my entire watchlist? No. Missing watchlist targets are logged, and usable models continue. Failed tag pages leave tags null while retaining the listing record. Persistent blocks produce a successful empty run with a status explaining the next step. Only malformed input fails validation.

Do I need an account on the target site? No. The Actor reads only public pages and never logs in.

How many models can I get? The Actor reads the public library in one page; its model count changes over time. Keyword searches start with site-ranked search results and can add matching library models. maxItems caps the final output, and filters may reduce the count.

Why is pullsNumeric approximate? The site shows pull counts in an abbreviated form such as 120M or 3.5K. The Actor keeps that text in pulls and converts it to a number in pullsNumeric, which therefore has the same precision as the abbreviation.

Why are capabilities or sizes empty for some models? Those lists mirror the tags shown on the listing. Models without such tags get an empty list.

Does it get community models uploaded by users? No. It covers the official library listing and its search results, and each record points to a library page.

What happens when the site changes its layout? The Actor parses the page by its structure, but a large redesign could break it. If you notice missing fields, please report it through the Issues tab so it can be fixed.

What if I get blocked? Retries with backoff are built in. On a blocked response, the Actor retries through fresh Apify Proxy sessions. You can also set a proxy configuration in the input.

Does it use AI enrichment? No. All values come directly from the page, with no third-party AI processor involved.

Is it legal to scrape Ollama?

The Actor collects only publicly visible information that anyone can see without logging in, and it does not touch private data or accounts. The model names, descriptions and statistics are public facts about software. Still, laws differ between countries, and a website's terms of service may restrict automated access. You are responsible for using the data in line with the site's terms and with the laws that apply to you, including copyright and data protection rules. The Actor spaces public page requests at least one second apart and backs off after retryable errors. If you are unsure about your use case, get legal advice.

Feedback

Found a bug, a missing field or have an idea for a new one? Open an issue from the Issues tab of the Actor page. Feedback with an example input and the expected output helps the most.

Search results are matched with the public library cards to include tagCount, updated and updatedAt when those cards are available. The sourceUrl identifies the library page for these matched records. If the library request fails, search records retain null values for metadata absent from search cards.