Your Claude subscription includes free cloud computers. Every cloud agent you start at claude.ai/code gets a fresh VM that keeps working after you close your laptop.
The problem: every new VM boots knowing nothing. It doesn't know which repos belong together, how they relate, what your conventions are, or what the last agent did.
BlitzOS fixes the problem by building you a context repo: one private GitHub repo owned by you that teaches every new cloud agent your whole setup:
your-context-repo/
CLAUDE.md # the map: your repos, how they relate, conventions, work loop
.gitmodules # member repos pinned by reference: no code is ever copied in
sessions/ # every cloud agent commits a record of what it did
docs/ # optional power-mode setup
Launching a cloud agent that can do real work becomes one click: the VM boots with the context repo plus every work repo auto-selected. Cloud agents read CLAUDE.md, read what previous cloud agents did, and start working with no blockers. Your laptop can be off the entire time.
Browser (fastest): sign in with GitHub at blitzos.com, pick the repos that belong together, and BlitzOS creates the context repo and gives you one-click launches, plus a feed that tracks every cloud agent's live status.
Local (most private): this skill builds everything on your machine and grants nothing to BlitzOS servers.
Self-host (yours end to end): run the same portal on your own Cloudflare account — your OAuth app, your database, your domain. One script, about 10 minutes: portal/SELF-HOSTING.md.
Needs Claude Code, Git, gh (authenticated), jq, and Node 18+.
git clone https://github.com/blitzdotdev/blitzos.git
cd blitzos && ./install.sh
claude "set up blitzos"Context repository, but for your whole company. Give cloud agents a role and launch them with scoped context, your team's tools (Google Workspace, Slack, Linear, Stripe, etc), and credentials. Agents become as capable as your coworkers.
What you get:
- Roles, not setups — give a cloud agent a role and launch: it gets the company repo (code, plans, research, docs, goals), your team's tools (Google Workspace, Slack, Linear, analytics, Stripe, etc) and the credentials that role needs.
- Whole loops, not just code — read the support ticket, check the code, ship the fix, reply to the customer, close the ticket from one cloud agent, start to finish.
- Policy layer over capability and credentials — every grant is per role and per agent, never account-wide; each cloud agent is a fork of your company holding only what its job needs.
- Audit trail end to end — everything every agent touched, when, and with which credential.
- Runs where you need it — our managed infrastructure or your own VPC.
- One-click launch a link that selects the context repo and every member repo, in the right order, every time.
- Warm multi-repo context cloud agents start knowing your conventions, the repo map, and what previous agents did.
- Native repository access Claude reads, writes, pushes branches, and opens PRs on every selected repo through Anthropic's own GitHub rail. Zero tokens in the default path.
- A cloud agent feed blitzos.com shows every cloud agent's live status (working / quiet / done) with a link to open any of them, from any device. One 2-minute environment setup.
- Your connectors cloud agents can use Linear, Slack, Gmail, and whatever else is already connected to your claude.ai account.
- A shared work log each cloud agent records PRs and decisions in
sessions/, so work compounds instead of evaporating.
The default flow needs zero credentials. The launch link selects the context repo and every member repo through Anthropic's native GitHub proxy, so all repository access stays on Claude's own rail. Your source code is never copied into the context repo (members are pinned by reference) and never touches BlitzOS servers.
Your agent setup should live in a repo, not on a machine.
Local Claude Code is powerful because your machine accumulates state: checkouts, skills, config, memory. That's also the trap: the state is stuck in one laptop. BlitzOS moves it into git, where it's versioned, reviewable, and portable. Once your context, your skills, and your work log live in a repo, a fresh cloud VM isn't a downgrade from your laptop — it's your setup, available in unlimited copies, from anywhere, in parallel, and always current. The durable unit of agentic work stops being the machine or the cloud agent. It's the context repo.
Shipped:
- Context repos — map + pinned member repos + sessions log, generated in one command
- One-click multi-repo launch — zero tokens, native GitHub rail
- Live cloud agent feed — cloud agents report status to blitzos.com; see working / quiet / done and open any cloud agent from any device
Shipping now:
- Clean-prompt status — cloud agents connect to the feed through a one-time environment key; nothing gets appended to your prompt
- Guided setup — the feed walks you through the 2-minute environment config and verifies it live
Next:
- Steer from the feed — reply to a running cloud agent from blitzos.com; the last reason to keep a terminal open goes away
- Skills travel with the repo — put your skills in the context repo and every VM boots with them: your
~/.claude, versioned and portable - Self-updating context — cloud agents propose PRs back to the context repo when they learn something (a new convention, a decision, a map change), so context compounds instead of rotting
- Codex support — the same context repo boots OpenAI Codex cloud agents; your context stops being vendor-locked
If a repo is missing in the default flow, confirm the Claude GitHub App can access that repository, then reopen the generated launch link. In power mode, follow the generated docs/CLOUD-SETUP.md for token setup and rotation.
MIT. See LICENSE.