Online communities often share reviews of using several AI tools side by side — ChatGPT, Gemini, and Claude among them. One question keeps coming up in those threads: what's the real difference between "just using several AIs in parallel" and "using one of them as a main agent — a central AI that gathers the others' outputs, compares them, and makes the final call"? After reading this, you'll understand that difference and be able to set it up yourself with just a Claude account. The basic click-through setup takes about 15 minutes.
What does it mean to use Claude as your "main agent"?
A main agent is the AI you put in charge of reviewing outputs or data from other sources, comparing them, and producing a final judgment. The key word is agent — an AI that plans and carries out work on its own, using tools and data, instead of just answering a single question. If you use several AIs separately, you have to compare ChatGPT's answer, Gemini's answer, and Claude's answer by hand every time. If you instead ask one AI to "gather what the others said, compare them, and tell me which has stronger evidence," that AI becomes the center, and the others become sources of raw material.
Anthropic's own published research-agent case follows a similar shape: a lead agent plans the overall task, breaks it into sub-tasks handled separately, and then pulls the results back together. The "main agent" workflow in this guide is the same idea, scaled down so a non-developer can replicate it with nothing but a Claude account.
What you need before you start
- A Claude account — the free tier is enough to follow the basic setup below.
- Outputs from the other AI tools you already use — save the answers you've gotten from ChatGPT, Gemini, etc. somewhere you can copy from.
- One topic to start with — pick a single task (planning a trip, comparing draft reports) rather than mixing several topics at once, so the setup doesn't get confusing.
Quick glossary
- Agent — an AI that plans and carries out work using tools and data, rather than just answering one prompt.
- Projects — a Claude feature that bundles custom instructions and reference files together, so every chat inside it shares the same context.
- Connector — a feature that lets Claude access outside services such as Gmail, Google Calendar, or Google Drive.
- MCP (Model Context Protocol) — a standard for connecting AI to outside tools — an open standard that lets AI applications connect to external tools and data. Anthropic describes it as a "USB-C port for AI applications": just as USB-C is one standard connector for many devices, MCP is one standard way to connect AI to many outside systems.
- Subagent — a helper AI that a main agent hands a specific task to, with its own separate workspace. This is a Claude Code (a developer tool) feature.
What non-developers can do right now: making Claude your main agent with Projects
You don't need to write any code to approximate a main-agent setup — Claude's Projects feature is enough. Follow the steps below.
- Create a new project — go to the project menu in Claude and create a new one. The exact button label can change with updates, so follow whatever the screen shows. Success looks like: a new project screen opens where you can name it.
- Write the role into the custom instructions — in the project's instructions field, spell out the role clearly, e.g. "In this project, review outputs from other AI tools and tell me which one has stronger supporting evidence." Success looks like: Claude follows that role in every chat inside the project, so you don't have to restate it each time.
- Upload the other AI's outputs as reference material — copy the answers you got from ChatGPT or Gemini and add them as text files, or paste them, into the project's file/knowledge area. Success looks like: the file you just added shows up in the project's file list.
- Ask it to compare and judge, in the chat — be specific: "Compare answer A and answer B — which has stronger evidence, and what's missing from each?" Success looks like: Claude refers to both pieces of material specifically and explains the differences and its reasoning. If it gives a vague answer that ignores your uploaded material, add "base your answer only on the uploaded material" to the instructions or the prompt and try again.
Letting it use your own data too, via Google connectors
Your main agent can look at your real data as well as other AIs' outputs. Claude can connect to Google Workspace services — Gmail, Google Calendar, and Google Drive — through connectors, available on Claude's web and desktop apps. Connecting requires you to authenticate with your own Google account, and on Team and Enterprise plans, an organization admin has to enable the connector before individual members can authenticate. Claude is designed to access your data only when a question or request actually calls for it, and to pull only the minimum information needed.
If you code: splitting work across subagents with Claude Code
If you're comfortable with code, you can go a step further. Claude Code, Anthropic's developer tool, supports subagents — helper agents the main agent can create and delegate specific tasks to. Each subagent gets its own workspace and its own scope of tool access, so you could split a large task into, say, "an agent that only does research" and "an agent that only writes code." Connecting Claude Code to outside tools such as a database or an internal API via MCP further extends what the main agent can reach on its own. Setting this up requires some familiarity with a development environment, so if it's your first time, check the official documentation first.
Common sticking points and fixes
| Symptom | Cause | Fix |
|---|---|---|
| You try to upload a file to a project and hit a storage-limit message | Free accounts have a smaller file capacity per project and fewer projects than paid plans | Delete old material you no longer need, check the exact limit shown on screen, and consider a paid plan if needed |
| The Google connector keeps disconnecting or won't authenticate | Your login session expired, or — on Team/Enterprise — the admin hasn't enabled the connector yet | Re-authenticate on a personal account, or ask your admin to enable the connector on an organization account |
| Asking it to compare answers produces a one-sided verdict with no real reasoning | It's judging only from what you uploaded, and vague source material can lead to a plausible-sounding but unsupported answer | Include the actual source links or full original answers, and have a person confirm any decision that matters |
Taking it further
Once you're comfortable with the basic setup, you can split projects by topic (a "trip planning" project, a "work report" project) and give each its own role instructions. If you code, the official Claude Code subagents documentation and MCP documentation cover more advanced automation. For more on Projects and connectors themselves, see the official Claude Projects help article and the Google Workspace connectors help article.
Frequently asked questions
Q. Do I have to use several AI tools at once?
No, there's no single right answer here. Claude alone is enough for plenty of tasks. But if you're already using several AI tools in parallel, giving Claude the job of consolidating and judging their outputs saves you from comparing them by hand every time.
Q. Can I do this on the free plan?
Yes, free accounts can use Projects, custom instructions, and file uploads. The number of projects and the file capacity are just smaller than on paid plans.
Q. Do I need to be a developer to use the "main agent" idea?
No. The Claude Projects approach covered here is entirely click-based, with no code required. Subagents that split up coding work, and MCP connections to outside tools, are features for developers using Claude Code.
Using several AI tools together isn't inherently right or wrong. But when their outputs stay scattered, the work of comparing them keeps piling up on you. Putting Claude in charge as the main agent — the one that consolidates and judges — hands off a good chunk of that work.