Every agent stack, pre-baked.
Each template is a Firecracker snapshot built from a Dockerfile or captured from a running sandbox, ready to restore in ~49ms. CPU and memory are baked into the template — pick one that fits or build your own. The vCPU count below is what the guest sees; sustained CPU depends on your plan, with Free capped at 2 cores per sandbox and Pro and Team bursting to all 8.
Base
--template base
Ubuntu 24.04 + mise with Node 22, Python 3.12, Go, and Bun pre-installed. The universal runtime that powers Apps.
- base
- ubuntu:24.04 + mise
- runtime
- node 22 · python 3.12 · go · bun
- cpu
- 8 vCPU burst
- memory
- 4 GB
- size
- 2.0 GB
- p50 boot
- ~49 ms
Code Interpreter
--template code-interpreter
Sandboxed Python + Node for data analysis, plotting, agent loops, and Functions. The default for LLM code execution.
- base
- python:3.11 + node 22
- runtime
- python 3.11 · node 22
- cpu
- 8 vCPU burst
- memory
- 2 GB
- size
- 12 GB
- p50 boot
- ~49 ms
Coding Agent
--template agent
One box with every terminal coding agent baked in — claude, codex, opencode, amp, grok, gemini, copilot. Drop the matching API key and start a long-running agent.
- base
- ubuntu:24.04 + node 24
- runtime
- claude · codex · opencode · amp · grok · gemini · copilot
- cpu
- 8 vCPU burst
- memory
- 2 GB
- size
- 6.0 GB
- p50 boot
- ~49 ms
Claude Managed Agents
--template claude-agent
Run Anthropic's Claude Managed Agents on PandaStack. The ant worker executes the agent's tool calls inside an isolated microVM — Claude drives the loop, your sandbox runs the code. Sub-second starts, stateful multi-turn.
- base
- ubuntu:24.04 + ant + mise
- runtime
- ant · node 22 · python 3.12
- cpu
- 8 vCPU burst
- memory
- 2 GB
- size
- 8.0 GB
- p50 boot
- ~49 ms
Browser
--template browser
Headless Chromium + Playwright + Xvfb, plus the Python crawl4ai extraction stack. Scrape, automate, screenshot, and extract.
- base
- ubuntu:24.04 + chromium
- runtime
- chromium · playwright · crawl4ai
- cpu
- 8 vCPU burst
- memory
- 4 GB
- size
- 4.0 GB
- p50 boot
- ~49 ms
PostgreSQL 16
--template postgres-16
Managed Postgres 16 with pgvector and pgbouncer on a durable volume. Backs the Databases feature.
- base
- ubuntu:24.04 + PGDG
- runtime
- postgresql 16 · pgvector
- cpu
- 8 vCPU burst
- memory
- 1 GB
- size
- 12 GB
- p50 boot
- ~49 ms
Custom build
--template custom
Bring a Dockerfile or build script. We bake the rootfs snapshot and serve it from your registry.
- base
- your image
- runtime
- anything that runs on linux
- cpu
- 8 vCPU burst
- memory
- you decide
- size
- you decide
- p50 boot
- depends on size
Live Snapshot
--template snapshot-template
Start from any running sandbox, install tools interactively, then save it with sandbox.snapshot({ name: "..." }).
- base
- running sandbox
- runtime
- whatever you installed
- cpu
- inherited
- memory
- inherited
- size
- captured state
- p50 boot
- ~49 ms
Bring a Dockerfile or snapshot a live sandbox.
Bake an image once into a Firecracker rootfs + snapshot, or install tools inside a running sandbox and save that exact state as a template.
FROM node:22-bookworm
RUN apt-get update && apt-get install -y --no-install-recommends \
git ripgrep fd-find tmux && rm -rf /var/lib/apt/lists/*
RUN npm i -g pnpm@9 && \
npm i -g @anthropic-ai/claude-code
WORKDIR /workspace
ENV NODE_ENV=production# Build the snapshot (~30s one-time bake)
# CPU and memory are baked into the template — set them here.
$ pandastack template build \
--name my-claude-stack \
--file Dockerfile \
--memory-mb 4096
# Restore a sandbox from it in ~49ms — it boots with 4 GB RAM (8 vCPU burst).
$ pandastack sandbox create \
--template my-claude-stack \
--ttl 1hOne-time bake
Build from Dockerfile layers or capture exact live sandbox state.
Instant restore
Every create restores from the baked snapshot — no cold boots after the first bake.
Reproducible
Snapshots are content-addressed. Pin a sha256 for production.
Don't see your stack?
Open an issue or send a PR — most community templates land within 48h.