Projects: coordination on top of sessions

What a project adds

A rebuilt Claude project adds a layer on top of single sessions. [Vendor-documented] "A project is one ongoing conversation where Claude coordinates a stream of related work for you." You say what needs doing. It starts a thread for each task.¹

A project does not change the architecture. It uses it. Each thread is a Claude Code session in its own sandbox. The project adds a coordinator that starts and tracks those sessions. It also adds standing context that every new session starts from. The Tour shows how to use a project.

Goals: why projects were rebuilt

[Vendor-documented] Anthropic names the problem: "Managing multiple sessions across a build used to require you to divide the work."² The rebuilt project has four goals:

Move the coordination from you to Claude. You "brief Claude in the project the way you'd brief a chief of staff."²
Set context once. Every new thread starts with the project's instructions. Every thread "draws from a shared memory."²
Keep the coordinator sharp. It "sees what threads report back, not every step they take."¹
Let work outlast one session. Cloud threads keep going after you close your laptop.¹

How it maps onto the four layers

Layer	In a project	Evidence
UI	The project conversation, thread cards and the Overview pane. They are at claude.ai/code, in the desktop Code tab and in the mobile app. We also started the new projects from chat in November 2026. Anthropic's page still calls chat projects "the earlier Projects experience"	[Vendor-documented], chat [Observed]
Coordinator	"One long-running session where Claude acts as coordinator." It has its own model and effort setting. It has no connectors	[Vendor-documented] Where it runs is [Unknown]
Threads	Each cloud thread is "a Claude Code cloud session working on its own branch and copy of the repo." It runs in the project's cloud environment	[Vendor-documented] The same kind of microVM is [Inferred]
Model	Threads run in the cloud or on your machine, "with Anthropic as the model provider in both cases". Projects are not offered through Amazon Bedrock, Google Cloud's Agent Platform or Microsoft Foundry	[Vendor-documented]

Sources: the Projects page and the launch post.¹,²

[Vendor-documented] A cloud thread loads each repository's CLAUDE.md. It also loads the skills, connectors and plugins on your account and project. It does not load a repository's own skills, .mcp.json or .claude/settings.json. Changes "reach new threads, not threads already running."¹

Where project state lives

[Vendor-documented] A project keeps three kinds of standing context. You set each one once.¹

Context	What it holds	Who reads it
Project memory	Notes Claude keeps, stored as files with a `MEMORY.md` index	Every cloud thread, at start
Project instructions	Up to 16,000 characters of standing rules	Each new thread and the coordinator
Repositories, files and environment	Cloned repositories, files under `/mnt/project-files`, the cloud environment	Every cloud thread

The Library tab holds the files you added and the files threads produced.

[Vendor-documented] The sandbox is still short-lived. "A cloud thread's sandbox pauses between turns and resumes when the thread continues." If it cannot resume, the thread starts "from a fresh clone, so uncommitted changes can be lost." Anthropic's advice is to "commit and push work in progress."¹

[Inferred] This is the rule from The sandbox and its state, built into the product.

Threads on your computer

[Vendor-documented] Some tasks need something only your computer has, such as a local database. Claude can run that thread on your machine through Remote Control. It uses that machine's files, tools, hooks and permission rules. It starts with the instructions "but not with its memory files loaded." It "runs only while that computer is awake."¹

Scheduled work inside a project

[Vendor-documented] Scheduled work in a project "runs as threads in that project." So each run starts with the project's instructions and memory.¹,³ One difference matters: "a thread that a scheduled task started doesn't wait" for a usage limit to reset. Its turn stops.¹ Scheduled tasks covers standalone tasks.

Authority in a project

[Vendor-documented] Threads run in auto mode when the model supports it. An approval prompt waits "inside that thread." Telling the coordinator to go ahead does not reach it. A rule such as "run at most two threads at a time" is saved to memory, but it is "not enforced settings." The enforced limits are your plan's limits and "200 new threads per day across your projects."¹ [Inferred] Words in a project do not grant or limit authority. See Security boundaries.

Sessions that message each other

[Vendor-documented] Claude Code has cross-session messaging for sessions you run yourself. One session can send another a plain-text message, such as a finding. It is "never the sender's conversation history or files." Messages between sessions on one machine stay on that machine. Messages to your cloud sessions or other machines go through Anthropic's servers. Claude finds those sessions only while the sending session is connected to Remote Control. Without that connection, the receiver cannot reply.⁴,⁵

[Vendor-documented] A message is information, not your voice. "It can't approve anything." "It can't change configuration." Commands in it never run. Permission prompts still fire in the receiving session. The crossSessionInbound setting can accept, hold or refuse messages. isolatePeerMachines makes Claude ask before a message leaves the machine.⁴

The worked example shows Brightline's payment run inside a project.

Why a project is still a personal tool

[Vendor-documented] "A project belongs to one user. You can't share a project or its threads with another user." There are "no organization-level controls for projects during the beta."¹ Projects are on Pro and Max plans. Anthropic says updated projects for Team and Enterprise "come after" the current rollout.² For several people giving work to Claude together, Anthropic points to Claude Tag.¹

[Inferred] So a project coordinates one person's AI Worker well. It is not the business layer from From personal to business worker. It has no worker identity, no business record you control and no enforced approval outside the thread.

[Unknown] ChatGPT projects hold files and context for Chat and Work, per the Tour. A documented coordinator with parallel threads was not found.

Rule to remember: A project coordinates sessions. It does not replace them. Each thread is a full session in a short-lived sandbox. Lasting state goes to memory, the Library or a pushed branch. Authority still lives in the service that acts.