Mikey builds Lumen, a liquid-glass journal app, in six Gemini 4 Argon prompts on Base44: tags and search, AI reflection, sign-up, Stripe premium and a final polish.
Run a fleet of cloud Claude Code sessions from one agent with Claude Projects
Chase shows how Claude Projects lets you talk to one agent that manages many always-on Claude Code sessions in the cloud, sharing context and memory, with a setup demo.
Chase AI · YouTube · published Oct 16, 2026 · Software & DevOps
▶ Watch on YouTubePrompt
Using Claude Projects, create a project for [my app] and have the lead agent split the work into [number] cloud sessions (frontend, backend, tests), share context between them, and report progress back to me.
Visual breakdown
Claude Code projects lets you talk to one coordinator agent that splits your work across a fleet of full, persistent Claude sessions running 24/7 in Anthropic's cloud, sharing context and memory and opening PRs in your GitHub repo. It draws from your Max plan usage, not API costs.
The problem: Normally you manage every terminal and agent yourself, keep them on the same page, and the work stops the moment your laptop shuts down.
How a Claude Code project runs your work
- Create the projectProjects beta in the desktop app: name, optional goal and context files
- Repo and environmentPick a GitHub repo for PRs; add a cloud environment with network access and secrets if needed
- CoordinatorOne orchestrator (Opus on low effort by default) splits the prompt and deals out work
- Parallel threadsFull Claude sessions with their own context, shared memory, each owning its PRsDropped: Blockers handed back to you for a decision or a key
- Unblock and steerAnswer blockers through the coordinator or talk to any thread directly
- Always-on projectPRs in your repo and a living build you can feed from anywhere, even with your computer off
Repo and secretsYou: setup and decisionsClaude sessionsResult
Use projects vs. a normal session
- Use it for complicated, long-running work you keep feeding and can split into parts
- Skip it for quick tasks you can finish in one chat
- Be wary when the work must touch files that live only on your machine
- Not on Team or Enterprise plans yet
By the numbers
5parallel threads from one prompt
25 minrun before the first review
< 24 hto get off the wait list
Set up your first project
Apply on the wait list page, then open Projects beta in the Claude desktop app.
- Archive old chat-era projects to start from a clean slate
- Click New project, give it a name, and optionally a goal and context files
- In settings, keep the coordinator on low effort and the thread model on Opus medium (or Sonnet)
- Keep project instructions short and turn on auto memory
- Select a GitHub repo, and add a cloud environment with secrets if you need API keys
- Allow local device access with a pre-approved folder only for work that must run on your machine
- Give it your prompt, then clear blockers through the coordinator
Turn on auto-continue so work resumes when your usage limit resets.
Tips from the demo
- The coordinator only deals out work, so low effort is fine; bump it up only if coordination is poor.
- Usage burns fast with 5, 10 or 15 threads working 24/7, so watch the thread model.
- Less is more in project instructions: every thread reads them.
- A session that needs your logged-in accounts (here, scanning X) can run locally and even build a local scheduled routine.
- Unlike subagents, threads persist, own their PRs, and you can message them at any time.
Jump to the part you need
0:00What Claude Code projects is0:41Why it matters: Claude as project manager2:24How to get access3:43When to use it and when not to5:24Projects vs. subagents6:52Demo: create a project7:47Settings: models, memory, environment12:23Giving it the AI Creator Radar prompt13:08Five threads after 25 minutes14:34Clearing blockers with secrets15:56The finished dashboard
Distilled from the video’s captions
How to try it
- Open Claude Code.
- Copy the prompt and fill in the [placeholders].
- Watch Chase AI's video for the full walkthrough and setup.
Related use cases
Leon shows cubefarm, which turns a coding agent swarm into a 3D office: Claude Code, Codex and OpenCode agents pick up GitHub issues from a Kanban board and open real pull requests.
Jack shows Replica, a free set of eleven Claude skills that reads an app's negative reviews, maps every screen, rebuilds it, tests each flow and fixes what users hate.
Riley shows eight everyday jobs Claude Code handles better than chat Claude: designing websites, building 3D models in Blender, editing videos and spinning up internal tools.