Claude Code Projects change how we brief long running client builds. Instead of juggling separate Claude Code sessions, you talk to one coordinator. Claude scopes the work, starts parallel cloud threads, and keeps a shared memory.

Anthropic describes the redesign as moving from a folder to a conversation. Official Claude Code docs and VentureBeat coverage from Carl Franzen (original launch around September 17, 2026) agree on the core model. This playbook is for India and USA agencies that ship multi repo client work.

We stick to verified product facts. No invented quotes. No invented limits. You get a delivery checklist you can share with coordinators and tech leads.

Claude Code Projects coordinator cloud threads and project memory diagram — Debugged Pro original diagram. Claude Code Projects snapshot for India and USA agencies.

Key Takeaways on Claude Code Projects

One coordinator conversation directs worker threads that do the coding work.
Each thread is a Claude Code cloud session on its own branch and repo copy, and can open PRs.
Shared project memory and a library carry decisions, files, and artifacts across threads.
Beta is on Pro and Max, rolling out gradually; Team and Enterprise come later.
Not in terminal CLI or IDE plugins yet; use claude.ai/code, desktop Code tab, or mobile.
Usage draws from your plan; multiple threads hit limits faster; docs cap 200 new threads per day.
Human PR review still matters; map oversight to NIST AI RMF style checks for client work.

What Anthropic redesigned in Claude Code Projects

Anthropic's Projects redesigned article says you describe what needs to get done and Claude manages the work. Claude scopes the request, delegates, coordinates parallel threads, reviews outputs, and assembles results. You can steer progress throughout, including from your phone.

Official docs frame a project as one ongoing conversation. You paste bugs, stack traces, or task lists into that conversation. Claude starts a thread for each piece or routes work to a thread already active in that area.

Two Anthropic examples help agencies brief clients fast:

Checkout p75 latency: profile endpoints, test optimizations, and open PRs in parallel threads.
Retire a deprecated v1 endpoint: connect API, web, and mobile repos; migrate callers; open PRs; learn merge order.

That multi repo pattern matches how India and USA delivery teams often staff client platforms. One goal. Several repos. One place to steer.

Claude Code Projects vs older Claude Projects naming

Names collide, so brief clients clearly. Older Claude Projects are a general workspace for chats, files, instructions, and knowledge. Each chat stays separate unless knowledge or memory captures context.

VentureBeat notes the September 17, 2026 launch coverage and this confusion. Claude Code Projects target agentic software development across multiple Claude Code sessions. Threads do the work. Claude directs it.

Existing chat Projects on Pro and Max keep working as they do today during rollout. Anthropic says the redesigned experience expands later across Claude, Team, and Enterprise.

How threads, memory, and Overview work

Under the hood, each thread is a Claude Code cloud session on its own branch and copy of the repo. The coordinator keeps work organized. If threads edit the same code, overlap becomes a normal merge conflict on the PR.

Threads can split work further with subagents, loops, and workflows. That helps large assignments finish faster without you becoming the human scheduler between sessions.

Claude Code Projects architecture with coordinator threads and Overview pane — Debugged Pro original diagram. How Claude Code Projects connect you, coordinator, threads, and Overview.

Shared project memory and library

Every thread adds to and draws from shared project memory. Anthropic examples include remembering a release moved to Friday, why an export was dropped, or who to check before touching billing.

Docs say project memory stores notes as files. Cloud threads read MEMORY.md when they start and open other memory files when needed. Project instructions (up to 16,000 characters) go to each new thread and to the coordinator.

The library collects files you add and artifacts Claude produces. Later threads can build on that pool instead of starting from a blank prompt.

Overview pane states agencies should watch

The Overview pane groups threads so you see what needs a human. Docs list groups such as:

Ready for review when a PR awaits review
Waiting on you when a reply, approval, or failure needs you
Working while a thread still runs
Landing when a PR is approved or queued to merge
Idle and Resolved for finished work

For overnight cloud work, this pane is the morning standup board. Open Ready for review and Waiting on you first. Then decide merges.

Beta access, GitHub setup, and where it runs

Docs state Projects are in public beta on Pro and Max and rolling out gradually. Early access favors accounts that used cloud sessions and do not already have projects in claude.ai chat or Cowork. Team and Enterprise are not available yet. Join the waitlist if the sidebar entry is missing.

Projects appear at claude.ai/code, in the desktop Code tab, and in the Claude mobile apps. They are not in the terminal CLI, VS Code extension, or JetBrains plugin yet. They are also not available through Amazon Bedrock, Google Cloud Agent Platform, or Microsoft Foundry per docs.

GitHub.com and the Claude GitHub App

Code projects need github.com (not GitHub Enterprise Server, GitLab, or Bitbucket for this path). Your connected GitHub account needs push access. The Claude GitHub App must be installed on the repos. A /web-setup token alone is not enough for project cloud threads.

That requirement matters for India USA agencies with mixed client hosting. If a client lives on GitLab or Bitbucket only, plan a different workflow until support expands.

Cloud by default, Remote Control when local tools matter

Cloud threads keep running after you close your laptop. When a task needs a local database, device emulator, or VPN only API, ask Claude to run that thread on your computer through Remote Control. Docs require Claude Code v2.1.280 or later on that machine.

Local threads use tools on your machine. They start with project instructions but not with project memory files loaded, per docs. They run only while that computer is awake with Remote Control on.

Usage limits and model settings for delivery teams

There is no separate Claude Code Projects price. Usage draws from your existing plan. Because each thread is a full Claude Code session, projects can reach limits faster.

Docs list an enforced limit of 200 new threads per day across your projects. Threads watching a PR can wake again when CI fails or review comments arrive. Idle projects with no running threads and no watched PRs do not burn plan usage while sitting quiet.

Set coordinator model and effort separately from thread model and effort in Project settings. A new project defaults to Opus everywhere, with high effort for threads and low effort for the conversation. For routine agency tasks, dial thread effort down after a few good dry runs.

Practical tip we use on Claude Dynamic Workflows for agencies: start with propose threads before auto start. Tighten instructions after one small real task. Then raise concurrency.

Agency playbook for India and USA teams

Debugged Pro delivery teams in India and the USA often coordinate migrations, latency goals, and multi repo client platforms. Claude Code Projects fit that shape when the goal outlasts one session. Here is how we would staff it.

Agency checklist for Claude Code Projects India USA delivery teams — Debugged Pro original checklist graphic for India and USA agencies using Claude Code Projects.

Pick work that belongs in a project

One goal across many repositories such as a shared lint standard or API version retirement.
An area you keep feeding with bugs and stack traces for one service.
A migration bigger than one session with decisions that must reach later threads.
Document heavy research where threads write files into the Library tab.

Skip a project for a single flaky test fix. Start a normal cloud session instead. Also skip when every task needs your machine only; use local sessions or agent view.

Standing instructions that protect client repos

Write project instructions before the first real batch. Include:

Which branch to start from and how to name draft PRs
Required checks such as make test and make lint before calling work done
Stop and report if a secret, connector, or repo is missing; do not guess
No merge, force push, or CI config change without human approval in the thread

Put repo specific build commands in each repo CLAUDE.md. Put cross repo decisions in project memory so later cloud threads inherit them.

Oversight checklist mapped to NIST AI RMF

Multi agent coding still needs human governance. The NIST AI Risk Management Framework gives agencies a shared language for map, measure, manage, and govern. For Claude Code Projects we translate that into delivery practice:

Govern: name who may create projects and who can merge PRs on client repos.
Map: list repos, secrets, connectors, and which tasks stay local via Remote Control.
Measure: track failed CI, merge conflicts, usage spikes, and Waiting on you aging.
Manage: require human review of every PR before merge; pause the project if a thread goes off policy.

India USA agencies can put that checklist into the SOW language. Clients care that overnight cloud threads do not merge themselves. Overview Ready for review is the control point.

Sample day for a delivery pod

Morning: open Overview. Clear Waiting on you. Review Ready for review PRs.
Midday: paste new tickets into the coordinator chat. Cap active threads if usage is high.
Evening: leave cloud threads running on profile or migrate tasks. Steer from mobile if needed.
Next morning: check Landing and Idle. Update project memory with any new client decision.

This rhythm fits distributed India USA pods. Cloud threads cover the overnight gap. Humans keep the merge gate.

What Claude Code Projects do not solve yet

No shared multi user project during beta; a project belongs to one user.
No CLI or IDE plugin surface for the coordinator experience yet.
Merge conflicts remain real when threads touch the same code.
Cloud sandbox pauses between turns; uncommitted work can be lost if resume fails, so commit and push on long tasks.
GitHub.com path only for the documented code project setup today.

Know these limits before you promise a client overnight autonomy. Sell coordinated assistance with a human merge gate. That is honest and still valuable.

How Debugged Pro can help

If your team wants a rollout plan for multi agent coding on client repos, start with our software development consulting conversation. We help India and USA agencies set instructions, review gates, and usage budgets that match real delivery calendars.

Browse more delivery notes on the Debugged Pro blog. Then pilot one Claude Code Projects stream on a non critical multi repo goal before you expand.

FAQ on Claude Code Projects

What are Claude Code Projects and how do they differ from older Claude Projects?

Claude Code Projects are a beta coordination layer in Claude Code. One coordinator conversation directs parallel worker threads that share project memory, instructions, and a library. Older Claude Projects are a chat knowledge workspace for grouping chats and files without that coordinator-and-threads model. Anthropic and VentureBeat both stress this naming difference.

Who can use Claude Code Projects in beta right now?

Official docs say Projects are in public beta on Pro and Max plans and rolling out gradually, starting with accounts that have used cloud sessions and do not have existing projects in claude.ai chat or Cowork. Team and Enterprise are not available yet. If you do not see Projects in the sidebar, join the waitlist. Threads are not available in the terminal CLI or IDE plugins yet.

Do Claude Code Projects threads open pull requests on GitHub?

Yes for code work on github.com. Each cloud thread is a Claude Code cloud session on its own branch and copy of the repo. Threads can open pull requests, watch CI with auto-fix, and report back. You need the Claude GitHub App installed and push access. Merge conflicts can still happen if threads touch the same code.

How does usage and the 200 threads per day limit work?

Projects draw from your existing Claude plan. Multiple threads use limits faster. Docs set an enforced limit of 200 new threads per day across your projects. You can set separate model and effort levels for the coordinator conversation and for worker threads. Check project Usage in settings to see token use by thread and model.

Can a thread run on my computer instead of the cloud?

Cloud is the default. When a task needs a local database, emulator, or VPN-only API, ask Claude to run that thread on your computer through Remote Control. Docs require Claude Code v2.1.280 or later on that machine. Other project threads can keep running in the cloud.

About the author

Nitin Kumar is Lead Project Coordinator at Debugged Pro. He works with India and USA agency teams on delivery process, multi repo coordination, and practical AI tooling. He focuses on clear checklists, human review gates, and client ready updates.