Product

Agent-first infrastructure

Stackap is an application platform that provides deployment, runtime, PostgreSQL, storage, domains, HTTPS, cron jobs, secrets, logs, metrics, monitoring, backups and rollback through an agent-operable control plane.

What it does

Most hosting products were designed for a person who clicks through a dashboard. The steps that cannot be done by a command, such as creating an account, finding a setting, copying a key from one screen to another, or reading a graph to decide what broke, are exactly the steps that stall a coding agent and send the work back to a human.

Agent-first means designing the other way round. Every operation exists as a command and as an HTTP call. Commands that list things can print JSON (projects --json today). Errors are written to say what to do next: a CLI that is not logged in prints the exact login command. The documentation is plain text an agent can read in one request. A person still decides what to build and approves what matters, but the repetitive operating work has a path that does not need them.

Four principles are used when deciding what to build. Do not send the human to a setting. Do not use the human as a courier between machines. Do not make the human diagnose something the agent can inspect. If the agent has to guess, the documentation has a bug.

How it works

Today the agent-facing surface is concrete and small:

• A command-line tool whose commands are one line each and take secrets from standard input, so a token or a variable never appears in a process listing. The CLI reference lists every command.
• An HTTP API with the same operations, authenticated by a token that belongs to exactly one organization. Scaling, settings and organization management are API calls.
• Logs, deployments and health are readable from the same tools, so an agent can inspect a failure instead of asking a person to describe it.
• Human approval for risky actions. With an approval-mode token, an action such as deleting a project returns approval_required; the account owner gets a decision link by a separate channel, approves or denies that exact request, and the agent then repeats it. Each approval is single-use and expires after an hour.
• A public description of the API at /v1/capabilities and /v1/openapi.json, with scopes and error codes, so an agent can learn the surface without a token.
• Plain-text descriptions of the site at /llms.txt and /llms-full.txt, generated from the same pages people read, so the two cannot disagree. See the agent documentation page.

An agent's whole path to production is a handful of commands, listed under AI app deployment.

In practice

$ stackap projects --json
$ stackap deploy my-app
$ stackap logs my-app --source build
$ stackap rollback my-app

Early-access scope

• There is no dedicated tool-protocol server for agents yet. An agent uses the command-line tool or the HTTP API.
• Some steps still need a person, by design or for now. Risky actions wait for an approval, creating an organization is an operator action, and pointing a domain at the server is a DNS edit made at the DNS provider.
• Documentation for agents is a single site summary and the CLI reference. A full, per-endpoint API reference has not been written.
• "Agent-operable" describes the design goal and the current commands. It is not a claim that any particular agent will operate it without mistakes, and every destructive step still deserves a human's approval.

Questions

Last updated 2026-10-03.