Deploy any code your agent ships.
Your agent writes the code. The Shed CLI deploys it.
What you install.
Two pieces, installed separately. The CLI deploys. The skill teaches your agent to drive it. Neither is open yet.
Setup guides, agent by agentThe CLIsoon
Packages your project, builds it on Shed, returns a live URL.
The agent skillsoon
Teaches Claude Code, Cursor, and Codex to write a SHED file and run the CLI.
MCP
A server your agent connects to instead of running the CLI.
Every host can run it. Count what it costs to get there.
Measured in handoffs, not minutes: the moments the agent stops and waits for you to open something. They are what actually make setup feel long.
- Google Cloud RunBilling and three APIs before the first deploy.
- create a Google account
- create a Cloud project
- attach a billing account
- install the gcloud CLI
- gcloud auth login opens a browser
- set the active project
- enable the Cloud Run API
- enable the Cloud Build API
- enable Artifact Registry
- answer the region and access prompts
- deploy
9handoffs
- VercelBuilt around a Git connection and a framework preset.
- create an account
- install the CLI
- vercel login
- link or create the project
- confirm the framework preset
- deploy
4handoffs
- RailwayThe deploy succeeds and nothing is reachable yet.
- create an account
- install the CLI
- railway login
- answer the railway init prompts
- railway up
- generate a domain in the dashboard
4handoffs
- Fly.ioOne launch command, but it interviews you.
- create an account
- install flyctl
- fly auth login
- add a payment method
- answer the launch prompts
- deploy
4handoffs
- Cloudflare WorkersNo billing form, but the scaffold interviews you.
- create a Cloudflare account
- install the Wrangler CLI
- answer the create-cloudflare prompts
- wrangler login opens a browser
- deploy, claiming a workers.dev subdomain
3handoffs
- ShedSign in once, then it is one file and two commands.
- create an account
- install the CLI
- shed login, approve in the browser
- shed init writes the SHED file
- shed deploy returns a URL
2handoffs
the agent can do it from the shell you have to go somewhere else
What you get, and what is coming.
Today
- Deploy from the CLIThe build runs on Shed, not on your laptop.
- Build and runtime logsStreamed live, or read after the fact.
- Status and historyEvery deployment and what happened to it.
- Cancel and retryStop a bad deploy before it replaces what is live.
- Scale to zeroIdle deployments cost nothing to keep around.
- Sandboxed by defaultA gVisor sandbox on a tenant-only node pool.
Next
- Private sharingsoonName the people who can open it. Nobody else can.
- Auth in front of your appsoonSign-in you do not have to build.
- Jobs and cronsoonRun it once, or on a schedule, and keep the exit code.
- More than one servicesoonSeveral pieces in one deployment, talking to each other.
- MonitoringsoonRequests, errors and latency without wiring anything up.
- AlertssoonTell you it broke before someone else does.
Before you hand it off.
What can I deploy with Shed?+
Web applications, APIs, workers, scheduled jobs, one-off scripts, or a container you bring yourself. Shed detects what the project needs and runs it.
Which coding agents work with Shed?+
Any agent that can run the Shed CLI. We ship a skill for Claude Code, Cursor, and Codex so they already know how.
How does Shed build and deploy my project?+
Shed builds from what is declared in the SHED file. Nothing outside it gets picked up: no ambient laptop state, no hidden dependencies.
Are deployments isolated?+
Yes. Each deployment runs in a gVisor sandbox on a node pool reserved for tenant workloads, so the kernel your code calls into is not the host's. Nothing shares a process tree with anything else you're running.
How does scaling work?+
A Shed scales to zero when idle and back up when a request arrives. The bounds are set by us for now; per-project limits are not exposed yet.
Why not Railway or Fly.io?+
Their unit is a project you tend: name it, add services, expose it in a second step. Agents produce software nobody tends, in volume. Different unit, not a worse cloud. Railway is a canvas.
Why not Vercel or Cloudflare?+
Vercel starts from a repository and a framework it recognises; an agent has a folder and twenty minutes. Cloudflare has every piece, and expects you to assemble them. Vercel is a workflow, Cloudflare is a kit of parts.
Where does my code run?+
Managed Kubernetes on GCP and AWS, each deployment in its own sandbox. Low-cost data centres are planned next, so idle software costs less to keep alive.
What do I get when a deployment finishes?+
Your agent receives a live URL in the same session. shed logs <id> streams stdout and stderr from any run.