Introduction

AgentServe hosts apps that coding agents build. An agent uploads a project directory and gets live URLs back, with no account and no sign-up step. The stack runs on a 1h TTL. If the human the agent works for wants to keep it, they open a claim link, create an account (or sign in), and the stack becomes theirs at the same URLs with its data intact.

A stack can have several services. A FastAPI backend and a React frontend that calls it is the typical case: you describe them in agentserve.yaml, and AgentServe builds both, wires the frontend's API URL to the backend, and starts them in dependency order.

The lifecycle#

Every anonymous stack follows the same path:

agent ──deploy──▶ live (1h TTL) ──expires──▶ snapshot kept 7 days ──claim──▶ restored, no TTL
                       │                                                  ▲
                       └──────── human opens the claim URL any time ──────┘

1. Deploy. The agent uploads its source as a tar.gz with the CLI, the MCP server or a plain HTTP request. It gets back the service URLs, a manage_token for later redeploys, and a claim_url to hand to its human.
2. Live. The stack serves traffic for its TTL, 1h by default. The agent can redeploy as often as it likes; the URLs don't change.
3. Expire. When the TTL runs out, the processes stop. Visitors see a "This preview expired" page. The built services and each service's $DATA_DIR are kept as a snapshot for 7 days.
4. Claim. Whenever the human opens the claim URL, during the TTL or during the 7 days after it, the stack moves into their account. A claimed stack has no TTL. If it had expired, it is restored from the snapshot without a rebuild, at the same URLs.

An unclaimed snapshot that nobody claims within the retention window is deleted, data included.

Who it's for#

• People working with coding agents (Claude Code, Cursor, Codex, Windsurf and others) who want to open what the agent built in a browser, share it, or keep it, without setting up hosting first.
• The agents themselves. Everything an agent needs is reachable without credentials: a CLI, an MCP server, a skill file served at /skill.md, and a small JSON API. Errors come back as text an agent can act on, and build and runtime logs are one call away.
• People who already have an account. Give your agent your API key and its deploys land straight in your account, with no claim step and no TTL.

What you get today#

• Python, Node and static services, up to 4 per stack, built and run as local processes.
• References between services (${api.url}) resolved at build time and at start-up.
• Build-then-swap redeploys with automatic rollback when the new version fails.
• Crash supervision with restart backoff, lifecycle events, and signed webhooks.
• A dashboard for claimed projects: a service map, deploy history, combined live logs, activity and settings.

Next steps#