What is Projects in Claude Code?

This guide is for people who use Claude Code for other work than building software, such as finance, analysis and documents.

Until recently, I typically worked in focused folders, running different agent tasks with different context, rules and setups. I started each session myself, told it what to do and kept an eye on all of them. I was the manager.

Now Anthropic has given us an assistant that does that coordination for us.

In a project in Claude Code, you write what needs doing, Claude starts a thread for each task, and the threads work in parallel. Most run in the cloud and keep going after you close your laptop.

Projects build into the app what we described in What is an AI agent?, one agent splitting a task and giving the pieces to other agents.

Anthropic launched projects on 17 September 2026 as a public beta for Pro and Max. They are on claude.ai/code, in the desktop app and in the mobile app, but not yet on Team or Enterprise. The gradual rollout starts with accounts that have used Claude Code in the cloud and have no projects in chat or Cowork.

A manager and its workers

Think of a project as a small team with a manager. Here it is in a finance team's project, with three threads doing very different work. One checks the month-end close, one writes the board report and one reviews supplier invoices.

• The conversation is the manager. It is where you send the work. Claude splits the job there, hands out the pieces and keeps track of them. A quick question is usually answered right there.
• The threads are the workers. Each one takes a single task and reports back to the manager when it is done.
• The Overview is the team board. It sits next to the conversation and shows the threads by status and the files they have made.

Three things make it easier to follow.

1. You send work in one place, the conversation. You do not start sessions yourself.
2. A thread asks for you when it needs an answer or an approval. Then you answer inside that thread.
3. Every thread in the cloud starts from the same setup, with the same instructions, files and connectors.

Each thread is a full Claude Code session with its own context window, the room Claude has to work in on one task. The manager sees what each thread reports back, not every step it takes.

When a project fits

A project pays off when the work has a goal that lasts longer than one session and keeps producing new tasks.

You do not need code to use one. Upload a folder of contracts or an export of supplier invoices you keep coming back to. Each thread saves its answer as a file in the Library tab.

Projects in Claude Code, chat and Cowork

Chat and Cowork also have projects, and they work differently.

A project in chat or Cowork is a workspace with its own instructions, files and memory. It suits thinking together, shared team context and work you want to steer turn by turn.

A project in Claude Code adds a manager that hands work to threads. It suits work that splits into parallel pieces and runs over days, and your job changes from doing each step to reviewing the results.

The instructions and memory of a Claude Code project are kept in the project settings on claude.ai, not in a folder of your own.

Keep your folder as the source of truth, as we describe in building AI capital, and copy the rules into the project. Anthropic writes that the new projects will also come to chat, Cowork, Team and Enterprise, without giving dates.

Projects and Claude Tag

Claude Tag is Claude in Slack. It is also in public beta, for Team and Enterprise. The difference is who the work belongs to.

Anthropic describes Claude Tag as the place for team work, and Cowork or Claude Code as the place for your own work.

The project instructions

A rule you write once applies to every thread after it.

If you edit the instructions, the threads that are already running keep the old version.

The project instructions work like the instruction file in an agent in a folder. Anthropic's documentation says good instructions cover five things.

1. What the project is for.
2. Where the work happens and what the results should be called.
3. How a thread checks its own work before it says it is done.
4. What a thread does when something it needs is missing.
5. What needs your yes before it happens.

Here are the instructions for the finance project.

YouCopy

This project handles the finance team's monthly work, one task per thread. Use the templates and last month's reports in Library. Before you say you are done, check every figure against the source file, and list the ones you could not match. If you lack something you need, such as a document, a login or a connector, say exactly what in your first message and stop. Do not guess, and do not fill the gaps yourself. Do not send anything out of the project before I have said yes in the thread.

The line about stopping comes from Anthropic's own example. Without it, a thread that lacks something may find a way around it instead of asking.

The Overview

The Overview sorts the threads into six groups by status. The one to watch is Waiting on you. It holds the threads that need an answer or an approval from you, and the ones that have failed. An approval sits inside the thread, so that is where you answer.

What it costs

Projects share the limits with the rest of Claude Code, but they use them up faster. Anthropic writes that on Pro in particular you should expect to reach the limit sooner on days you run a project.

A project with no threads running and no new messages from you uses nothing.

A thread that hits the limit waits until it resets and then carries on. Work you left running also uses your plan after the reset. Stop the thread or pause the project if you do not want that.

For a new task, ask Claude for a new thread, because an old thread that has been idle for over an hour rereads its whole history first.

Threads on your own computer

When a task needs a program, a database or files that live on your computer, you can run that one thread on your own machine.

Files can be uploaded to the project, but I still prefer working against files on my laptop. When everything sits there, it is easy to switch between AI agent tools and simply point them to the folders.

1. In the Claude desktop app, open Settings, then Claude Code, turn on Use this computer from your phone and claude.ai, and add the folder to the list below it.
2. Choose Work locally in the + menu in the project conversation, and allow the folder on the card that appears.

The thread runs only while your computer is awake and the desktop app is open. It usually works in that folder without asking you each time. It gets the project instructions and the setup on your machine, but not the notes Claude keeps for the project.

If every task needs your computer, a normal Claude Code session there fits better than a project.

GritAI comment

We still recommend building your agents in folders, so you can use other AI tools on the same folders. That said, your laptop has to be on if you want to work on them from your phone.

Larger agent teams with subagents

A thread can also hand parts of its task to helpers of its own, called subagents. In the finance team, the board report thread could have one helper for revenue, one for costs, one for cash and one that writes the commentary.

Claude Code has a few general helpers built in. Helpers with a set role, such as one for revenue, have to be set up first.

For threads in the cloud, they live as files in a code repository the project uses or in a plugin you add in the project settings. The helpers report to their thread, and the manager only sees what the thread passes on.

We show how to design a team with a lead and specialists in How to build a team of AI agents.

Your first project

Pick work you will come back to over the next few weeks. Upload the documents and write a few lines of instructions.

A new project runs every thread on Opus at high effort, and that uses the plan the fastest. The conversation itself starts at low effort, which is the setting you see next to the message box. Check Thread model and Thread effort in the project settings first. Then send one small real task with this line.

YouCopy

Propose the threads before you start them, and run at most two at a time until I say otherwise.

Open each thread when it is done to see what it did and how it reports. When the first threads deliver what you wanted, let Claude run more at once.

If you would rather learn it with others, we run Claude Code bootcamps in Oslo, where we work hands-on with AI agents all day.