How EasyCanDo stays up to date

Once a week, a computer looks for new tools and for changes to the 1,317 tools listed here. Then a person reads every change before it goes live.

Where the information goes · Tap any box for the details

14 sources

Only one thing skips approval: a re-check that finds no change, which only moves the tool’s checked date.
Read every step as a list
  1. Step 1 · code on a Mac mini, no AI

    Find what’s new since last Sunday

    Every Sunday at 3 a.m. Pacific, a Mac mini reads all 14 sources: about 600 downloads and API calls in all.

    After it reads a source, it saves what it saw. The next Sunday it reads the source again and sets the two side by side. Only the difference goes on.

    For example

    • Last Sunday OpenRouter listed 470 models. If it lists 474 today, the 4 new ones go on, and the 470 it already saw don’t.
    • A pricing page that showed $20 and now shows $25 goes on. One that shows the same amounts doesn’t.
    • Feeds such as Hacker News, X and the newsletters are only asked for posts since last Sunday.
    New source
    Its first Sunday only saves a copy, so adding a source never floods the list

    The 14 sources

    • Source · public API, no key

      OpenRouter

      The exact call

      GET https://openrouter.ai/api/v1/models

      Downloads OpenRouter’s full model list: about 470 models, each with its price per million tokens in and out.

      What counts as new

      • A model maker that wasn’t on the list last Sunday and isn’t in the catalog: a possible new tool.
      • A new model from a maker the catalog lists, such as a new Claude or Gemini model: a possible change.
      • A price that moved on a listed maker’s model: a possible change.
    • Source · public API, no key

      Vercel AI Gateway

      The exact call

      GET https://ai-gateway.vercel.sh/v1/models

      Downloads the list of models Vercel’s AI Gateway serves, about 500, each with its type: text, image, video or speech.

      What counts as new

      • A model maker the catalog doesn’t know: a possible new tool.
      • A new image, video, speech or transcription model from a listed maker: a possible change.
    • Source · public API, no key

      models.dev

      The exact call

      GET https://models.dev/api.json

      Downloads models.dev, an open list of about 320 model providers and their models.

      What counts as new

      • A new provider with at least one model of its own, not only other makers’ models resold.
      • At most 5 a week, the most models of their own first.
    • Source · public file, no key

      LiteLLM

      The exact call

      GET https://raw.githubusercontent.com/BerriAI/litellm/
          main/model_prices_and_context_window.json

      Downloads the price list the open-source LiteLLM project keeps on GitHub, and reads the providers in it, about 150.

      What counts as new

      • A provider that’s new to the file and unknown to the catalog.
    • Source · public API, no key

      MCP Registry

      The exact call

      GET https://registry.modelcontextprotocol.io/v0.1/servers
          ?limit=100&updated_since=<last Sunday>
      (page by page, at most 40 pages)

      Reads the official registry of MCP servers, the plug-ins that let AI assistants use a tool. Only servers added or updated since last Sunday.

      What counts as new

      • A listed tool that published its own server: a possible change.
      • A company’s first server that runs online, if the catalog doesn’t know the company: a possible new tool. At most 10 a week.
      • Servers under personal GitHub accounts are skipped.
    • Source · public API, no key

      Hacker News

      The exact call

      GET https://hn.algolia.com/api/v1/search_by_date
          ?tags=story&query=Launch HN
          &numericFilters=created_at_i><last Sunday>
      GET https://hn.algolia.com/api/v1/search_by_date
          ?tags=show_hn
          &numericFilters=created_at_i><last Sunday>,points>=40

      Two searches of Hacker News through its public search: every Launch HN post (mostly new startups), and Show HN posts with 40 or more points, since last Sunday.

      What counts as new

      • A Show HN post must link to a site and use words like AI, API, model, agent or MCP.
      • Tools already listed are dropped. At most 15 a week.
    • Source · GitHub’s API, signed in

      GitHub releases

      The exact call

      gh api repos/<owner>/<repo>/releases/latest
      (once for each of 91 repos, 4 at a time)

      Asks GitHub for the latest release of each listed tool that keeps its code there.

      What counts as new

      • Nothing goes on to be checked. A new release is written to the engine’s change log, for the record.
    • Source · public API, no key

      Hugging Face

      The exact call

      GET https://huggingface.co/api/models?sort=trendingScore&limit=100
      GET https://huggingface.co/api/spaces?sort=trendingScore&limit=100

      Reads the 100 trending models and the 100 trending apps (Spaces) on Hugging Face.

      What counts as new

      • Created since last Sunday, with 20 or more likes, from an owner the catalog doesn’t know.
      • Copies and noise are skipped: compressed versions of other models, adult content, workflow files.
      • At most 10, the most liked first.
    • Source · X’s official API, paid

      X

      The exact call

      GET https://api.x.com/2/tweets/search/recent
        query = (launching OR launched OR introducing
                 OR "now available" OR "open source"
                 OR "open-sourced" OR "just shipped")
                (AI OR API OR MCP OR agent OR agents
                 OR model OR SDK OR LLM)
                has:links -is:retweet -is:reply lang:en
        start_time  = last Sunday (X keeps 7 days)
        sort_order  = relevancy
        max_results = up to 100 a page
        tweet.fields = created_at,public_metrics,entities

      Searches the past week of posts on X: original posts in English that announce a launch and carry a link.

      What counts as new

      • Only posts with 10 or more likes.
      • Posts are grouped by the site they link to, and ranked by likes plus twice the reposts.
      • Links to tools already listed, and to X itself, are dropped. At most 15.
      Cost
      $0.005 per post read, stopping at $10 a month
      Kept
      The tool’s link and the post’s id, never the post’s text
    • Source · public RSS feeds

      TLDR AI, AINews, The Rundown AI

      The exact call

      GET https://tldr.tech/api/rss/ai
      GET https://www.latent.space/feed
      GET https://www.therundown.ai/feed
      then GET each new issue’s page

      Reads the three newsletters’ feeds, and every issue since last Sunday, up to 14 a newsletter.

      What counts as new

      • A story that says something launched, released or opened up, and links to a site the catalog doesn’t know.
      • Sponsored stories are skipped.
      • The three take turns, so one long newsletter can’t fill the list. At most 15.
    • Source · 425 pages on the tools’ own websites

      Pricing pages

      The exact call

      GET <each listed tool’s pricing page>
      (425 pages, 6 at a time)

      Downloads the pricing page of every notable listed tool and keeps a short fingerprint of it: every dollar amount on the page, and any warning words such as deprecated, sunset, shutting down or waitlist.

      What counts as new

      • A dollar amount that appeared or disappeared since last Sunday: the price may have changed.
      • Warning words that appeared.
      • A page that’s gone: not found two Sundays in a row.
    • Source · 43 pages on the tools’ own websites

      Changelogs

      The exact call

      GET <each listed tool’s changelog or release notes>
      (43 pages)

      Downloads the changelogs of notable listed tools and looks only for warning words.

      What counts as new

      • Warning words that appeared since last Sunday, such as deprecated, discontinued or shutting down.
    • Source · an OpenAI model with web search

      Web scout

      The exact call

      POST https://api.openai.com/v1/responses
        model     = gpt-6.1-terra
        tools     = web search, up to 20 searches and page opens
        1 call a week

      An AI model searches the open web for launches the other sources missed: launch posts, changelogs, Product Hunt, GitHub, Hugging Face and news.

      What it’s told

      • “Search the web for ones that launched, or opened to the public, between <last Sunday> and <today>. Only things a person can sign up for, call or download now.”
      • It gets the name of every tool the catalog lists, to leave out.
      • It must name the page it read for each one.

      What counts as new

      • Up to 15, each with its name, site, launch date and one plain sentence.
      • Tools the catalog knows are dropped again, in code.
    • Source · the form on this site

      Visitor suggestions

      The exact call

      The suggest form on /about
      → an email to the person who runs the site
      → forwarded into the engine

      When a visitor suggests a tool or reports a mistake, the person who runs the site reads it and forwards the ones worth checking.

      What counts as new

      • Each forwarded suggestion becomes one lead, marked unverified.
      • A reported mistake about a listed tool becomes a warning on that tool.
      • Suggestions go first in line among possible new tools.
  2. Step 2 · code, no AI

    Already in the catalog?

    Each new item is matched against the catalog by its name and its website.

    Then

    • Not in the catalog: a possible new tool.
    • In the catalog: a possible change to that tool, such as a new price, a new model, warning words on its page, or a page that’s gone.
    • The same tool found by two sources counts once.
    • Anything already checked, or already waiting for approval, is skipped.
  3. Step 3 · code, no AI

    Choose this week’s checks

    Each check in step 4 is part of an AI call with web searches, and that costs money, so the engine checks at most 120 leads a week. It fills the 120 in this order:

    The 120

    • 30 possible new tools, best sources first: visitor suggestions, then X, the newsletters and Hacker News.
    • 30 routine re-checks: listed tools that have gone longest without a check, even though no source flagged them. That keeps every listing’s checked date fresh.
    • 60 of the most urgent of the rest: a page that’s gone, warning words, a price change, then new MCP servers, new tools and new models.
    • In a quiet week, room left over goes to more re-checks.

    What doesn’t fit

    • It goes on the waiting list. Next Sunday it competes for the 120 again, ahead of new leads of the same kind.
    • If more arrives each week than fits, the list grows. After 8 weeks unchecked, a lead is dropped, and the list holds at most 300.

    Two spending limits

    • The checks stop early if the next call could take this run past $20, or this month past $100. Whatever wasn’t checked goes on the waiting list.
  4. Step 4 · OpenAI, gpt-6.1-terra

    AI checks each one

    The exact call

    POST https://api.openai.com/v1/responses
      model     = gpt-6.1-terra, medium reasoning
      tools     = web search only
                  up to 40 searches and page opens a call
      store     = false
      up to 24 calls a week, 5 leads in each

    For each lead, the model opens the tool’s own pages (pricing, docs, changelog, repo) and says what’s true, with links to what it read.

    What it’s told

    • “Your knowledge is months stale: check every claim against a primary source with web search, opening the page itself, before you propose anything.”
    • “Research only: never sign up, enter keys, pay, or install or run anything.”
    • Short lessons from past approvals and rejections.

    Its answer for each lead

    • No change, with proof on the tool’s own site: the tool is marked checked today. No approval needed, since nothing a visitor reads changes.
    • Skip, when it’s noise, minor or can’t be confirmed: dropped.
    • Update, retire or new tool, with the new details: on to step 5.
  5. Step 5 · code and OpenAI

    Prepare the listing

    Turns each change into the listing you’d see on the site.

    In order

    • Plain words: OpenAI rewrites the catalog’s notes into the sentences on the site. 20 listings a call, no tools, at most $2 a week. Wording a person fixed by hand is never rewritten.
    • Pictures: for a tool whose output you can see, one download of the maker’s own page, then one OpenAI call picks the pictures. The maker’s robots.txt is honored.
    • A sample answer: for a data or text API, its docs are read and one OpenAI call picks a real example to show.
    • A last check, in code: a listing with banned words, a research note or a missing field is held back.
    Then
    The finished listing goes on a private Approve page, and the person gets a text
  6. Step 6 · a person

    Human approval

    A person opens the private Approve page. Each new, changed or leaving tool shows as it will look on the site, before and after. They approve or reject each one.

    Then

    • Approved: saved to the catalog.
    • Rejected: it never goes live, and the site keeps what it had.
    • If a tool changes again before it’s published, it comes back for another look.
  7. Step 7 · when the person presses Republish

    Publish

    The Mac mini rewrites the site’s data files from the catalog, letting through only approved listings. No AI calls.

    Then

    • One commit to the site’s code on GitHub, named for the week.
    • Vercel sees the commit and rebuilds easycando.com.
    • Each tool page shows the date it was last checked. The weekly update lists tools added and retired, and changes to price, access, abilities or name.

Where things are kept

  • Kept on the Mac mini

    Last Sunday’s copies

    One file per source, holding what it saw last Sunday: OpenRouter’s model list and prices, each pricing page’s dollar amounts and warning words, which posts and issues were already read. Step 1 replaces them every Sunday.

  • Kept on the Mac mini

    The waiting list

    Leads that didn’t fit in this week’s 120. Next Sunday they compete again, ahead of new leads of the same kind. It holds at most 300, and a lead is dropped after 8 weeks unchecked.

  • Read in step 4

    The open web

    The tools’ own pages: pricing, docs, changelogs and code repos. The AI only reads them. It never signs up, pays or installs anything.

  • Kept on the Mac mini

    The catalog

    The engine’s master list of every tool: what it does, how to use it, its prices, and the day it was last checked. Only approved changes go into it. The site shows its public part: 1,298 tools.

  • What you read

    easycando.com

    Rebuilt by Vercel after each Republish, from approved listings only.

See this week’s update

See a mistake, or a tool that’s missing? Tell the site owner