Deploying from an AI assistant

Tinycloud is built for the case where the person who wants the app is not the person who deploys software. The MCP server is the surface an assistant drives.

pnpm mcp

It is a JSON-RPC/MCP stdio server. It uses TINY_API_URL / TINY_TOKEN, or the current CLI profile when those are absent.

The CLI and the MCP server both use the same remote REST API. This keeps an agent and a human from getting different answers to the same request.

The loop that works

Probe the target before generating anything

tiny_get_platform_capabilities

This is the step that matters most. It reports whether the target executes build commands, offers Postgres, allows direct egress, and what memory ceiling applies. Generate a manifest the target can actually run rather than discovering it during a deploy.

Validate the manifest

tiny_validate_manifest

Parses, normalizes, and applies semantic and policy rules. No mutation, no upload.

Plan

tiny_plan_deployment

Returns the resource diff, migration plan, capability status, required approvals, and a digest that makes a stale plan detectable. Still non-mutating.

Deploy, then poll

tiny_deploy_app  →  tiny_get_deployment

Deploy is asynchronous. A returned deploy is not a finished deploy.

Read logs on failure

tiny_get_logs

Errors are results, not transport failures

Tool errors are returned as structured tool results with stable codes, remediation, retryability, and next actions.

{
  "status": "approval_required",
  "reasons": [{ "code": "NEW_PRODUCTION_CAPABILITY", "detail": "stripe.payments.read" }],
  "next_actions": ["tiny_approve_deployment"]
}

An agent can act on next_actions directly. It never has to parse prose to decide what to do next.

What an agent can do without a human

Ungated, because none of it widens blast radius:

redeploy an app whose permissions did not change;
create and delete previews — including with capabilities, which previews do not gate;
read status, logs, environments, access, audit events, and usage;
roll back.

Gated, because it does:

a new production capability, raw secret, outbound host, widened visibility, or destructive migration. See Approvals.

Writing manifests an assistant can deploy

Prefer capabilities over secrets and egress

A capability is the only outbound path that needs no approval in a preview and leaves nothing in the VM to leak.

Prefer SQLite unless the app needs concurrent writers

It needs no operator configuration at all, so it deploys on a minimally configured target.

Skip build.command when you can

Plain TypeScript with no build step deploys on a target that never installed the build image. @tinycloud/runtime-sdk is supplied by the runtime.

Use relative URLs for browser assets

On a single-host install, apps live below /apps/<route>/. An absolute /styles.css resolves above the prefix and 404s.

Full tool list

See the MCP reference.

Get started

Build apps

Operate

Deploying from an AI assistant

The loop that works

Errors are results, not transport failures

What an agent can do without a human

Writing manifests an assistant can deploy

Full tool list

​The loop that works

​Errors are results, not transport failures

​What an agent can do without a human

​Writing manifests an assistant can deploy

​Full tool list

The loop that works

Errors are results, not transport failures

What an agent can do without a human

Writing manifests an assistant can deploy

Full tool list