Claude Projects (redesigned)

#claude-code#anthropic#agent-orchestration#cloud-sessions#coding-agents#coordinator#threads#beta

Claude Projects (redesigned)

A project is one ongoing conversation where Claude coordinates a stream of related work, starting a parallel cloud-session "thread" per task instead of you managing several sessions yourself. Public beta on Pro/Max, announced Sep 17, 2026.

Projects Redesigned: From Folder to Conversation

What changed

The old "Projects" in claude.ai chat and Cowork was a folder: conversations plus reference files, no coordinator. The redesign makes a project one conversation where Claude itself decides what becomes a thread, delegates to it, reviews what comes back, and assembles results. Each thread is usually a full cloud session — Claude Code running in the cloud on its own branch — that keeps going after you close your laptop. When a task needs something only your machine has (a local DB, a VPN-only API), you can ask Claude to run that one thread on your computer instead, through Remote Control.

Why it matters here

This is Anthropic's own official entry into the category moc-agent-workbenches tracks — and it changes the shape of the question in Is there a free, open-source Cursor alternative?. That post's gap analysis was: per-agent attention signals exist (cmux, Superset), and per-task graph tracking exists (Beads), but nothing joins them into one coordinator that decides what needs a new thread, watches all of them, and tells you what's ready. Projects does roughly that natively, just scoped to Claude Code alone and without Beads — closer to what bb's plugin model was speculated to eventually reach.

How it's organized

  • The project conversation: a long-running coordinator session. It routes messages — quick questions get answered in place, new work becomes a thread or joins an existing one, several unrelated asks in one message become separate threads.
  • Threads: the workers, each a separate session with its own context window. A cloud thread works its own branch and opens a PR when the work calls for it, then watches it with auto-fix on.
  • Project memory: notes Claude keeps as files (MEMORY.md index + detail files) — decisions, pitfalls, requirements — read by every new thread. Separate from a repo's CLAUDE.md, which threads also load.
  • Project instructions: up to 16,000 characters sent to every new thread — which branch to target, how to self-check, what needs a go-ahead first.
  • Overview pane: threads grouped as Ready for review / Waiting on you / Working / Landing / Idle / Resolved, plus Library (files in and out), Pull requests, and Routines tabs.

Coordination is steerable, not configured

There's no settings panel for concurrency or update cadence — you tell Claude in plain language ("run at most two threads at a time", "propose threads and wait for my go-ahead", "only post when something finishes or is blocked") and it saves the preference to project memory. It's a followed convention, not an enforced cap.

Notable constraints

  • Beta: Pro/Max only, rolling out gradually to accounts that already used cloud sessions; Team/Enterprise later.
  • GitHub only (github.com, not Enterprise Server/GitLab/Bitbucket), via the Claude GitHub App, not just a personal token.
  • A project belongs to one user — no sharing a project or its threads, no org-level controls yet.
  • Draws on the same plan limits as any other Claude Code session, just faster, since several threads can run at once (enforced cap: 200 new threads/day across your projects).
  • Not available in the terminal CLI, Bedrock, Vertex, or Foundry.
Edit this page

Links to this page:

  1. MOC: Agent Workbenches