Docs

Getting started — stub

Runtimes are in private preview. This page is the shape of the docs, not a live API reference. Nothing here will provision a machine.

Persistent, policy-controlled Python execution for AI agents.

Request access at hello@pysmith.com. When an instance exists, you will get a connector URL and a SKU. Until then, treat the snippets as intent.

1. Connect MCP

Agents discover tools over the Model Context Protocol. Point your client at the instance connector when we issue one. Local stdio wrappers may follow; HTTP is the default.

{
  "mcpServers": {
    "pysmith": {
      "url": "https://api.pysmith.com/mcp/<instance>"
    }
  }
}

Expected tools, names not final: exec_python, list_files, spawn_agent. Your planner should not need SSH.

2. Install packages

The instance is a real interpreter. Use uv or Pixi. Installs land in the persistent environment and stay there. We keep a per-tenant cache so the second pandas install is not a blank CDN trip.

# on the instance, once
uv pip install pandas httpx numpy

# later, from an agent
exec_python("import pandas as pd; print(pd.__version__)")

We will not mount a shared read-only site-packages across customers. If a wheel is on your desk, it is yours.

3. Persistence

The workspace directory survives process restart and overnight idle. Child processes you start stay running until you stop them or the instance is deleted. That is the product. It is also why the SKU is billed while the instance exists.

# Tuesday
open("notes.txt", "w").write("still here")

# Friday
open("notes.txt").read()  # 'still here'

Limits (preview)

CPU and RAM are the SKU. No silent bursting.
Isolation is a system container, not a microVM.
No GPU SKU on this page, because we are not shipping one.
Quotas, regions, and SLAs: unpublished. We will not invent them.

Docs

When the stub becomes a reference.

If you already know the SKU you want, skip the prose and request access.

Request access See pricing