Ship AI Agents
With Confidence.
An open-source TypeScript runtime designed for production. Isolated sandboxes, multi-turn orchestration, and streaming capabilities out of the box. Positioned perfectly between raw LLM SDKs and bloated frameworks.
Everything You Need to Ship Agents
From sandbox isolation to multi-agent orchestration, Sandcaster handles the infrastructure so you can focus on your agent's logic.
Isolated Execution
Run untrusted AI code securely. Built-in adapters for E2B, Docker, Vercel, and Cloudflare for fresh sandboxes per run.
Streaming API (SSE)
Native Server-Sent Events support. Stream tokens, tool calls, and state changes to the client with zero latency overhead.
Speculative Branching
Fork agent states safely. Test different tool paths simultaneously and merge the most successful result back to main.
Multi-Agent Orchestration
Define supervisor agents that delegate tasks to specialized sub-agents with strict type-checked communication channels.
Multiple Interfaces
Don't just build APIs. Sandcaster ships with ready-to-use hooks for CLI, TUI, HTTP APIs, and native Slack bot integration.
Config-Driven via JSON
Manage environments, model selection, and telemetry settings globally using a simple sandcaster.json configuration file.
Built for the Terminal & Beyond.
Sandcaster's core philosophy is decoupling logic from the interface. Build your agent once, deploy it as a script, a persistent API, or an interactive terminal application.
Installing dependencies...
Generating sandcaster.json...