Fatsearch AI agents tokens is a search and read layer for AI agents. Its repo describes searching and reading the web for AI agents at 99% fewer tokens, with a free CLI and MCP server. For a social agent, that means trend checks, source reads and analytics context cost far less, so a reply workflow stays cheap to run.
Your social agent (Claude, ChatGPT, Cursor...)
|
|-- fatsearch MCP / CLI --> web search + read
| (small, clean context)
|
'-- Creator OS MCP --> posts, comments, DMs,
mcp.creatoros.ca/mcp analytics, Skool, ads
|
v
Instagram TikTok YouTube X Facebook LinkedIn
Threads Reddit Discord Slack WhatsApp SkoolWhat Does Fatsearch Do for AI Agents and Tokens?
Fatsearch gives an agent two verbs, search and read, and returns compact answers instead of full pages.
That is the whole idea. A normal research step in an agent loop looks like this: search, fetch three or four pages, paste raw HTML into the context window, then ask the model to reason over it. Most of those tokens are navigation, cookie banners, scripts and repeated boilerplate. You pay for all of it on every turn, and you pay again when the conversation grows.
Fatsearch sits in front of that. Your agent asks a question, gets back a short result set or a cleaned read of one page, and moves on. The project describes the result as search and read for AI agents at 99% fewer tokens, shipped as a free CLI and MCP server. The source is public: github.com/fatsearch/fatsearch.
The point is not that the web gets smaller. It is that your agent stops carrying the parts of the web it never needed. For social work, the thing you actually need is usually one number, one quote or one headline plus a link. Everything else is cost.
How Do Social Agents Waste Tokens on Research and Replies?
Most of the waste in a social agent is not the writing. It is the gathering before the writing.
A realistic daily loop for a social agent has three expensive moments:
- Trend check. The agent searches a topic, opens four pages, and summarizes. Raw pages are the biggest single line item in the context.
- Reply drafting. The agent pulls the comment thread, the original post text, and sometimes your past replies to match tone. Comment threads repeat the same author names and timestamps on every call.
- Analytics reading. The agent fetches per-post analytics and daily metrics, then re-reads them on the next turn because it did not keep a short summary.
None of these need full pages. The trend check needs a sentence and a link. The reply draft needs the comment and your voice, not the whole thread formatted as JSON with every field. The analytics read needs the numbers you asked about.
On the Creator OS side, the same principle applies inside the account actions. There are about 160 MCP tools, and the specific ones you want are narrow: list_comments, reply_to_comment, get_post_analytics, get_best_time_to_post. Calling get_post_analytics for one post returns less than pulling account-wide metrics and filtering in your head. The social agents repo is the harness that runs this loop, and the Social Agents repo post and the social agents docs cover how it is wired.
How To Wire the Fatsearch CLI and MCP Into a Social Reply Agent
Add fatsearch as a second MCP server next to Creator OS and the agent picks the cheap read path on its own.
Follow the setup in the fatsearch repo to install the CLI and start its MCP server. Then give the same agent two servers. For Claude Code, the Creator OS side is one command:
claude mcp add --transport http creatoros \
https://mcp.creatoros.ca/mcp \
--header "Authorization: Bearer cos_live_..."Fatsearch goes in the same MCP config file next to it, with whatever command and arguments the repo lists. Cursor, Windsurf and VS Code use an mcpServers entry with type http for Creator OS and the matching entry for fatsearch. Claude and ChatGPT use the hosted connector flow, described at the MCP docs.
Then write the instruction that makes the split explicit. A system prompt like this works:
You run my social accounts.
For any fact, stat, headline or link, use fatsearch search
and read. Do not fetch full pages.
For anything on my accounts, use Creator OS tools:
list_comments, reply_to_comment, get_post_analytics.
Keep at most 3 sources in context. Summarize each in
one line plus its URL, then drop the raw text.Two details make it stick. First, tell the agent to summarize and drop, not summarize and keep. Second, keep the account tools read-scoped until you trust the loop. A read-only Creator OS connection only sees read tools, and the API refuses writes from it, so a draft-only week costs nothing in risk.
When you want the reply loop to run without you, the CLI is the faster path than prompting each time:
creatoros inbox:comments --platform instagram --limit 20
creatoros inbox:reply post_123 --accountId acc_456 \
--message "Thanks, the file is in the guide."
creatoros automations:create \
--accountId acc_456 --keywords "guide,link" \
--dmMessage "Here it is." --commentReply "Sent to your DMs."That is the comment-to-DM pattern from the use case page. The keyword matching happens in Creator OS, so the model never has to read the thread to decide who gets the DM.
Token-Savings Math for a Daily Posting and Reply Workflow
Run the numbers on one day and the difference is easy to see.
Here is a simple model. Assume a typical raw page read costs a few thousand tokens once you include markup, and a cleaned read costs a fraction of that. The exact ratio is the number the fatsearch project advertises at about 99% fewer tokens. Your own mileage depends on the pages.
| Step, once per day | Naive loop | With fatsearch |
|---|---|---|
| Trend research, 4 pages | 4 full page reads | 4 short reads plus links |
| Reply drafting, 20 comments | Full thread JSON, twice | Comment list, replies only |
| Analytics context | Account-wide metrics | One post, one metric call |
| Carried in context next turn | Everything above | Three one-line summaries |
The multiplier that bites is the last row. Raw page text does not disappear after the agent uses it. It stays in the conversation and gets re-sent with every following turn. So a 4 page research step at the start of a 20 turn reply session is not paid once, it is paid 20 times. Cutting the page text by roughly 99% cuts that recurring block by roughly the same amount, because the block is what recurs.
This is not an argument for thinner output. It is an argument for keeping the raw web out of the conversation and keeping only the sentence you will use. The context window optimization post goes deeper on what to keep and what to drop.
Fatsearch vs TokenForge-Style Context Optimization
They solve adjacent problems: fatsearch cuts what enters the context, and TokenForge-style tooling compresses what is already there.
A TokenForge-style approach treats context optimization as a pass over your text: summarize, deduplicate, trim history, compress instructions. It is useful, and it applies to any input, including your own notes and drafts.
Fatsearch changes the input itself. If the raw page never enters the conversation, there is nothing left to compress. The two stack well: fatsearch first so the page stays out, then a compression pass on the conversation history if it still grows.
| Fatsearch | TokenForge-style optimization | |
|---|---|---|
| Where it acts | At the source, before the read | On text already in context |
| Shape | CLI and MCP server for search and read | A compression or summarization step |
| Best for | Research, trend checks, source reads | Long threads, accumulated history |
| Works with Creator OS via | Second MCP server in the same agent | Prompt or middleware step |
Pick based on where your tokens go. If most of your context is fetched web pages, cut the fetch. If most of it is your own back and forth over 40 turns, compress the history. Most social agents have both problems and want both tools.
Which AI Agent for Social Media Fits a Cheap Context Layer?
Any MCP capable agent works, because both layers speak MCP.
Creator OS does not ship its own model. It connects to the AI app you already use: Claude, ChatGPT, Cursor, Claude Code, Codex, Windsurf and VS Code are listed in the docs. Because the hosted server is a plain MCP endpoint with an auth header, adding fatsearch next to it is a config change, not a rebuild.
The split looks like this in practice:
- Fatsearch answers “what is happening out there” cheaply.
- Creator OS answers “what are my accounts doing and what do I publish or reply”.
If you want the loop packaged with brand context and a local dashboard, Social Agents is the open-source harness: start with npm start creatoros social-agents, it interviews you about your brand, then posts, replies, runs automations and reports. It lives at github.com/kevinbadi/social-agents, and the open-source Social Agents post walks through the harness. For a faster read on the setup, watch the walkthroughs on the KevBuildsApps YouTube channel and the launch video.
Can a 99% Token Cut Improve Your AI Reporting Agents?
Yes, and reporting is where the cheap read layer pays off fastest.
A weekly report agent has a fixed shape: pull analytics, look up context for anything unusual, write a short summary. Step two is where reports get expensive, because the agent wanders the web to explain a spike and drags pages back with it.
Wire it the other way. Have Creator OS produce the numbers with narrow calls, then let fatsearch handle only the outside context:
creatoros analytics:daily --from 2026-01-01 --to 2026-01-07 \
--platform instagram
creatoros analytics:best-time --platform tiktok
creatoros analytics:posts --sortBy engagement --limit 10Then the prompt is short: “explain the two biggest changes, use fatsearch for any outside fact, one source per claim, no full page reads.” The report comes back with three lines of research instead of three pages, and the numbers came from your own accounts rather than a guessed figure.
The same logic applies to replies in your voice: keep your own examples and tone notes in the prompt, then let the agent read one comment at a time through list_comments rather than re-reading the whole thread each time you reply. The token side is simply not carrying the thread twice.
FAQ
How many tokens does fatsearch actually save per search?
The project describes its search and read layer as returning results at 99% fewer tokens, which is a claim about the size of what comes back compared with reading raw pages. The real saving on your workflow also depends on how many turns carry that text forward, because fetched page text is re-sent with each later turn. Measure one day of your agent’s context before and after to get your own number.
Is the fatsearch CLI and MCP free to use for social agents?
The fatsearch repository describes the CLI and MCP server as free, and it is an open project you can read before installing. Your other costs are separate: the AI app you run the agent in, and Creator OS if you want it to act on your accounts. Creator OS includes API keys, the MCP server, the CLI and the agent skills in every plan.
How do I connect fatsearch MCP to my AI social media agent?
Install it from the fatsearch repo, start its MCP server, and add it to the same MCP config your Creator OS connection uses. For Claude Code the Creator OS entry is claude mcp add --transport http creatoros https://mcp.creatoros.ca/mcp --header "Authorization: Bearer cos_live_...", and fatsearch sits beside it. Cursor, Windsurf and VS Code use an mcpServers entry for each.
Does cutting tokens hurt reply quality or context accuracy?
Not if you cut the right thing. Removing page markup, cookie banners and repeated thread fields does not change the comment you are answering or your tone. Removing the comment itself would. Keep the primary object, the comment, the post text and your voice examples, and drop the packaging around them.
Fatsearch vs TokenForge: which is better for social agents?
They do different jobs, so the question is which problem dominates your context. Fatsearch stops raw pages from entering, and TokenForge-style optimization shrinks what is already in the conversation. A social agent that does daily trend research and analytics reads usually benefits most from the first, then adds the second if long sessions still grow.
Get started
Connect both servers to one agent and keep the raw web out of your reply loop. Creator OS plans start at $14.99/month for 11 platforms plus Skool, with the API, the MCP server and the CLI included. Start at https://www.creatoros.ca/sign-up, then read the MCP docs to add the connector and pick read only while you test.
Get started for $0
Free trial for new subscribers. Cancel anytime.