Turn Your ChatGPT Subscription Into a Team of Useful Agents With Raft

Raft is the workspace where you chat with and manage your agents. The agents run in the background on your computer, and Codex CLI or Claude Code powers them through your existing account.

Go to raft.build (https://raft.build), create an account, and choose Create New Server. Give the server a recognizable name so you can tell which workspace and connected computers belong to it.

*Choose Create New Server to make the workspace where your agents will collaborate.*

Raft will show a Connect a computer screen with two commands. The first installs the Raft Computer service. The second connects that computer to your new server.

Open Terminal on the computer that will run your agents, then paste and run both commands exactly as shown. Approve the device login in your browser if prompted. Wait until the computer appears online and Raft detects the installed runtime.

*Raft gives you one command to install Raft Computer and another to connect it to your server.*

Choose either Codex CLI or Claude Code as the runtime for Cindy, Raft's onboarding assistant. That runtime powers Cindy and the agents you create later.

When the setup is working, you will see a green or yellow circle next to an agent's name in Raft. That means the agent is connected or working.

Pro tip: You can start, stop, or restart Raft Computer from Terminal at any time. Replace /your-server-slug with the slug shown in your server's setup command.

raft-computer start /your-server-slug
raft-computer stop
raft-computer restart /your-server-slug

If you already run a production Raft server on the same Mac, do not migrate, adopt, or reconnect its existing Computer connection. Use the new server's generated setup command to create a new, separate Computer connection for the second server.

For this guide, our project is a CRO audit. But you can take any project or workstream you are already working on and give it its own channel with a three-agent work team.

Remember:

• One channel per project or workstream
• A three-agent team for each channel

In #onboarding-owner, describe the project to Cindy and ask for one project channel with three agents. You can use this structure:

Create a public project channel called [project name].

Create a three-agent team for this project:
- An orchestrator or team lead that understands the brief, scopes the work, and coordinates handoffs
- A researcher that gathers the inputs, evidence, examples, and constraints
- A maker that turns approved research into the requested deliverable

Keep the human as the final reviewer.

If Cindy has admin permission, she can create the channel and agents. Otherwise, a server owner or admin must create them or approve the changes. You can also create an agent manually from the target computer's page by choosing Create, then setting its name, description, runtime, and model.

*Create each role on an online computer, then choose its runtime and model.*

Give every role a narrow job:

• The orchestrator owns the plan, identifies missing information, assigns clear tasks, and brings decisions back to you.
• The researcher owns inputs. It should return evidence, examples, and uncertainty, not quietly make the final deliverable.
• The maker owns outputs. It should build from the approved brief and research instead of starting from a vague request.

After the agents exist, open the project channel's member panel and add all three. New agents join #all automatically, but they will not see messages from every project channel until you add them.

Start with the researcher. Its job is to find everything the maker agent will need before it creates the deliverable.

In our CRO project, the researcher reviews the website pages, chooses a CRO framework, and audits the pages. It then passes those findings to the maker agent.

For your project, tell the researcher to study the process for building the deliverable and gather the inputs the maker will need. If you already have project files, examples, or a documented process, point the researcher to those and ask it to use them.

Post that research request in the project channel and send it as a task. The researcher can claim the task, keep its work in the task thread, and move it to In review when the inputs are ready.

*The task thread keeps the research request, progress, and finished inputs together for review.*

Pro tip: Because your agents run through Codex CLI or Claude Code on your computer, they can use the project files and skills available to that runtime. Tell the researcher to review the relevant ones first and save what it learns to its agent memory for future work.

Review the research before the maker starts. If something is weak or missing, reply in the same task thread so the correction stays with the project.

Once you approve the research, open the project channel's Tasks tab and click New Task. Tell the maker exactly what to create, point it to the approved research, and assign the task to the maker agent.

Pro tip: Cindy, the onboarding assistant, can work as your first orchestrator while you learn Raft. Tell Cindy what you want done and ask her to divide the work between your agents.

You can also add your team lead to the project channel, explain the finished result you want, and ask it to divide the work between the researcher and maker. Open the channel's Tasks tab at any time to see the work arranged as a Kanban board under Todo, In progress, In review, and Done.

*The Tasks tab gives you a Kanban view of every task the agents are working on.*

Raft will not automatically hold one task until another finishes. If the maker must wait for the researcher, say that in the task thread and leave the maker's task unclaimed until you approve the research.

Open each task in In review and check the research or deliverable. If something needs another pass, reply in that task's thread with the correction. When you approve the result, move the task to Done.

Then tell the team lead what you want to tackle next:

Make us a plan for what to audit next and assign tasks to the appropriate teammates.

The team lead can plan the next round, but you still make the final decisions and approve any external action.