A smaller tool list
Apps like GitHub come with dozens of tools, each with its own instructions. Behind ATR, your AI sees 7 small tools and looks up the one it needs: 917 tokens instead of 32,563 for GitHub.
A little less overhead. A lot more room.
Less Tokens, More Usage
ATR is a free connector for Claude, ChatGPT and other AI apps. It searches the web, reads long documents and filters large tool results so your AI sees the answer, not thousands of unnecessary tokens.
Start using ATRFree & open source · No account needed
One question.
Two very different payloads.
Ask which of 30 GitHub issues are about OAuth. Without ATR, your AI reads every issue and field. With ATR, it gets the two relevant issues and can retrieve anything omitted later.
ATR / Token overview
Example · GitHub issues
“Which open issues are about OAuth?”
Measured · GitHub · 30 issuesWithout ATR
26,235 tokensEvery issue title, body, label, author and metadata field enters the conversation.
With ATR
420 tokensOnly the issues that answer the question enter the conversation. The full result stays retrievable.
The matching issues kept. The other 28 left out, but saved if asked for.98% fewer tokens in this result ↗
Measured November 6, 2026 on GitHub's official MCP server. Savings vary by task. See all results ↓
At home with the AI tools you already use
ClaudeChatGPTCodexCursorand other MCP clients
More signal.
Less baggage.
AI apps measure reading in tokens, roughly ¾ of a word each. Everything your AI reads counts: your messages, its replies, and every web page or tool result it pulls in. It re-reads all of it with every new step. A single list from GitHub can be 26,000 tokens, which uses up a chat's room and your plan's limits fast.
ATR sits between your AI and the places it reads from. It passes along the parts that answer your question, quoted exactly, and keeps the rest on hand in case your AI asks for it.
Explore what ATR doesYour AI appClaude, ChatGPT, Codex…
ATRFind → Run → Reduce
Your tools & dataWeb · Documents · MCP
Only the part you asked about
Your AI tells ATR what it is looking for. ATR returns the matching sentences, rows or lines, word for word, from the live page or app.
Nothing is lost
The full result is kept for a while. If your AI needs a detail that was left out, it can search or filter it without loading everything again.
What it's good for
ATR helps most when your AI would otherwise read a lot to answer a little. Here's what that looks like in a chat.
Questions about your GitHub (or other connected apps)
“Which open issues are about login bugs?”
Instead of every issue in full, your AI gets the few that match.
26,235 → 420 tokensReviewing pull requests
“Any open PRs that touch the payments code?”
Only the matching pull requests come back, not 30 full descriptions.
28,284 → 392 tokensPages that change
“What's on the Hacker News front page right now?”
ATR reads the live page. In our tests, Claude's built-in reader returned an old copy every time.
Right every time in our testsExact numbers and versions
“What Python version does this library need?”
ATR quotes the page word for word, so numbers aren't paraphrased or mixed up with an older version.
Quoted, not summarizedOne question about a long document
“In this PDF, what scores did the model get?”
Works with PDFs, Word, PowerPoint and Excel links. You get the matching sentences instead of 40 pages.
About 250 tokens per answerCoding with an AI agent
“Run the tests and fix what fails.”
With ATR installed on your computer, test runs keep the failures and drop hundreds of “passed” lines.
6,685 → 675 tokensMeasured results
We tested ATR in real claude.ai chats and against GitHub's official MCP server on November 6, 2026. Tokens are estimated as characters ÷ 4. The comparison includes where ATR saves tokens, where it improves freshness, and where it costs more.
| What your AI has to read | Without ATR | With ATR |
|---|---|---|
| Instructions for every tool (in apps that load them all) | 32,563 | 917 |
| 30 open issues: “which are about OAuth?” | 26,235 | 420 |
| 30 open pull requests: “which touch streamable HTTP?” | 28,284 | 392 |
| Repository search: “which have the most stars?” | 4,232 | 369 |
| A README: “what's the install command?” | 1,160 | 213 |
| Claude's built-in reader | ATR | |
|---|---|---|
| Right answers | Wrong on every Hacker News and README question (0 of 8) | 28 of 28 |
| Tokens for all four tasks | 1,698 | 4,144 |
Big savings on app data
When your AI pulls lists, issues, records or files from a connected app, ATR cut what it had to read by 82–99%.
Fresher, exact answers from the web
ATR read the live page every time. Claude's built-in reader is cheaper on simple pages (about 2.4×), but it served old copies of pages that change.
You may need to ask for it
In claude.ai, Claude uses its own reader unless told otherwise. One line in your preferences fixes that (see setup below).
Set it up in two minutes
A connector is an add-on that gives your chatbot new tools. Copy this link, add it as a connector in your app, and ATR is ready in your next chat.
https://atr-mcp.vercel.app/mcp
No account or sign-in. Choose “No authentication” when asked.
Claude & Claude Desktop
- Open Settings → Connectors → Add custom connector.
- Name it ATR, paste the connector link, and click Add.
- Open a new chat. Make sure ATR is switched on under + → Connectors.
- Claude will use its built-in reader unless you ask for ATR. To make ATR the default, add this line to Settings → Profile → personal preferences (or a Project's instructions): “When I ask you to search the web or read a link, use atr_search and atr_fetch instead of the built-in web tools.” Or skip the setup and ask your question and add “use atr”, like “What is the latest on the MCP spec? Use atr.”
ChatGPT
- Open Settings → Apps & Connectors → Advanced and enable Developer mode.
- Choose Create, name it ATR, and paste the link.
- Set Authentication to No authentication.
Other MCP apps
Add the link as a remote MCP server using Streamable HTTP. Available connector options depend on your app and plan.
Local coding tools
For local files, shell commands and coding tools, install ATR on your computer:
pip install atr-mcp
atr setup
What your chatbot gets
Seven small tools. You don't need to learn them: your AI picks the right one. These are the names you'll see in the chat.
Read the web & documents
atr_fetch reads a link (web page, PDF, Word, PowerPoint, Excel or JSON) and returns the sentences that answer the question.
Search the web
atr_search asks DuckDuckGo, Bing and Brave at once and returns one short list: title, link and one line each. Five results are about 300 tokens.
Inspect a website
atr_site shows a site's title, description, language, headings and link counts, or searches its public sitemap and navigation.
Find local places
atr_places finds cafes, restaurants, stores, services and other places using free OpenStreetMap data. No API key needed.
Check the weather
atr_weather returns current conditions and up to seven forecast days for a city, postal code or address.
Look up what was left out
atr_query searches or filters a full result ATR kept, so your AI can get one more detail without reading everything again.
See what you saved
atr_stats shows how many tokens ATR saved in your chat. Ask “how much did ATR save?”
Put your other apps behind ATR
This is where the biggest savings are. If you use an online MCP server that takes an API key or needs no sign-in (GitHub, DeepWiki and many others), add it here. You get one personal connector link. Use it instead of the original connector, and your AI gets the same tools with much smaller results.
Not yet supported: apps you sign into through your chatbot, like Canva, Notion, Figma or Gmail.
Your personal link contains encrypted credentials. Keep it private: anyone with the link can use those servers.
Servers that need you to sign in through a browser (OAuth) can't go behind ATR yet. Keep those as their own connectors.
Less to store.
Less to worry about.
ATR keeps no database and doesn't record what you read. Large results are held in memory briefly for follow-up queries, then dropped.
Where are my server keys stored?
In your personal link, encrypted. They aren't stored in a database. Treat the link like a password and only add it to apps you trust.
Can ATR work with email, Word, PowerPoint or Excel?
atr_fetch can read Word, PowerPoint and Excel files from a public link. Sending email or editing cloud files requires access to your account, so those actions need a connected email or Microsoft MCP server. The public ATR connector never asks for those credentials.
Can the hosted connector read my local files?
No. It reads public websites and online documents. It cannot reach your computer, local files or private networks.
Does ATR change my usage limits?
ATR reduces tool overhead. It doesn't change your plan's limits, compress your conversation history or shorten the AI's own replies. Savings depend on the tools and tasks you use.
When doesn't ATR help?
Short, simple web pages in claude.ai: Claude's built-in reader uses fewer tokens there. Also short answers from apps, where there's little to trim. ATR helps most with long pages, documents, pages that change, and large results from connected apps.
What about tools that make changes?
Read operations and changes use separate tools. Changes go through atr_call so your chatbot can request approval before running them.