# upstream

> Per-compute-second agent sandboxes on bare-metal Firecracker microVMs. Self-hosted, multi-model, ConnectRPC + REST + WS terminal. Anonymous demo sandboxes, BYOK keys, per-session budgets.

## Docs

- [Quickstart and concepts](https://docs.upstream.build/concepts.html): sandboxes, templates, sessions, auth
- [REST + ConnectRPC API reference](https://docs.upstream.build/api.html): every endpoint, request shape, response shape
- [Python SDK](https://docs.upstream.build/python-sdk.html): `upstm-py` — sync + async clients, `Upstream`, `Sandbox`, exec, files, ports
- [OpenAI Agents integration](https://docs.upstream.build/openai-agents.html): `openai-agents[upstream]` provider — drop-in compute for the Agents SDK
- [Templates](https://docs.upstream.build/templates.html): `claude-code`, `python`, `nodejs`, `base` — VM-ready rootfs specs
- [Pricing](https://docs.upstream.build/pricing.html): per-vCPU-second + per-GiB-second, no egress tax

## Optional

- [Anonymous demo](https://upstream.build/demo): browser-native Claude Code session, no signup, no API key — one per IP per hour
- [Landing](https://upstream.build/): positioning, hero, pricing summary
- [Console](https://console.upstream.build/): dashboard, API keys, session history