Introducing Sandcaster

Why We Built Sandcaster

Most developers building AI agents face the same frustration: raw LLM SDKs are too low-level, while existing frameworks are too opinionated. You end up either writing a lot of boilerplate plumbing, or fighting an abstraction that doesn’t fit your use case.

We wanted a runtime that sits at the right level of abstraction — one that handles sandbox lifecycle, streaming, multi-model routing, and agent orchestration out of the box, without locking you into a particular framework or cloud provider.

That’s Sandcaster.

What It Does

Sandcaster is a TypeScript runtime for AI agents that run in isolated sandboxes. It handles the infrastructure so you can focus on agent behavior.

• Isolated execution: Every agent run is isolated in its own sandbox — E2B, Docker, Vercel, or Cloudflare Workers. No shared state, no cross-contamination.
• Multi-model: Switch between Anthropic, OpenAI, Google, and OpenRouter without changing your agent code.
• Streaming by default: All responses stream over SSE. The API, SDK, and CLI all support streaming out of the box.
• Config-driven: Define your agents in sandcaster.json. No code required for common configurations.

Key Features

Starters — Six built-in agent templates you can deploy immediately: general-assistant, research-brief, document-analyst, support-triage, api-extractor, and security-audit.

Multi-agent orchestration — Run multiple agents in parallel or sequence. Pass context between agents using the composite sandbox model.

Skills system — Extend any agent with reusable capabilities defined in SKILL.md files. Skills are composable and version-controlled alongside your code.

SDK and API — @sandcaster/sdk provides a typed TypeScript client. The Hono-based API server handles routing, auth, and streaming.

CLI with TUI — The Ink-based CLI gives you a rich terminal interface for running agents, watching logs, and inspecting sandbox state.

Slack integration — The included Slack bot lets your team trigger agents directly from Slack channels.

Getting Started

bun add @sandcaster/sdk

Create a sandcaster.json in your project root:

{
  "starter": "general-assistant",
  "model": "claude-sonnet-4-5",
  "sandbox": "e2b"
}

Run your first agent:

bunx sandcaster run "Summarize the latest commits in this repo"

What’s Next

We’re actively working on:

• Search integration — Built-in web search and document retrieval for research agents
• Analytics — Run metrics, cost tracking, and latency dashboards
• Interactive API playground — Test and debug agents directly from the browser

Get Involved

Sandcaster is open source. Read the docs to get started, or explore the source on GitHub.

We’re building this in public. If you run into issues or have ideas, open an issue — we read everything.