Sandbox Agent SDK

Universal API for automatic coding agents in sandboxes. Supprots Claude Code, Codex, OpenCode, and Amp.

• Any coding agent: Universal API to interact with all agents with full feature coverage
• Server or SDK mode: Run as an HTTP server or with the TypeScript SDK
• Universal session schema: Universal schema to store agent transcripts
• Supports your sandbox provider: Daytona, E2B, Vercel Sandboxes, and more
• Lightweight, portable Rust binary: Install anywhere with 1 curl command
• OpenAPI spec: Versioned API schema tracked in docs/openapi.json

Roadmap:

[ ] Python SDK [ ] Automatic MCP & skillfile configuration

Agent Support

Feature	Claude Code*	Codex	OpenCode	Amp
Stability	Stable	Stable	Experimental	Experimental
Text Messages	✓	✓	✓	✓
Tool Calls	—*	✓	✓	✓
Tool Results	—*	✓	✓	✓
Questions (HITL)	—*		✓
Permissions (HITL)	—*		✓
Images		✓	✓
File Attachments		✓	✓
Session Lifecycle		✓	✓
Error Events		✓	✓	✓
Reasoning/Thinking		✓
Command Execution		✓
File Changes		✓
MCP Tools		✓
Streaming Deltas		✓	✓

• Claude headless CLI does not natively support tool calls/results or HITL questions/permissions yet; these are WIP.

Want support for another agent? Open an issue to request it.

Architecture

• TODO
  • Local
  • Remote/Sandboxed

Components

• Server: TODO
• SDK: TODO
• Inspector: inspect.sandboxagent.dev
• CLI: TODO

Quickstart

SDK

• Local
• Remote/Sandboxed

Docs

Server

• Install server
  • curl (fastest & does not require npm)
  • npm i -g (slower)
  • npx (for quick runs)
• Run server
• Auth

Docs

CLI

Docs

Tip: Extracting API Keys

TODO: npx command to get API keys

Project Goals

This project aims to solve 3 problems with agents:

• Universal Agent API: Claude Code, Codex, Amp, and OpenCode all have put a lot of work in to the agent scaffold. Each have respective pros and cons and need to be easy to be swapped between.
• Agent Transcript: Maintaining agent transcripts is difficult since the agent manages its own sessions. This provides a simpler way to read and retrieve agent transcripts in your system.
• Agents In Sandboxes: There are many complications with running agents inside of sandbox providers. This lets you run a simple curl command to spawn an HTTP server for using any agent from within the sandbox.

Features out of scope:

• Storage of sessions on disk: Sessions are already stored by the respective coding agents on disk. It's assumed that the consumer is streaming data from this machine to an external storage, such as Postgres, ClickHouse, or Rivet.
• Direct LLM wrappers: Use the Vercel AI SDK if you want to implement your own agent from scratch.
• Git Repo Management: Just use git commands or the features provided by your sandbox provider of choice.
• Sandbox Provider API: Sandbox providers have many nuanced differences in their API, it does not make sense for us to try to provide a custom layer. Instead, we opt to provide guides that let you integrate this project with sandbox providers.

FAQ

Why not use PTY?

PTY-based approaches require parsing terminal escape sequences and dealing with interactive prompts.

The agents we support all have machine-readable output modes (JSONL, HTTP APIs) that provide structured events, making integration more reliable.

Why not use features that already exist on sandbox provider APIs?

Sandbox providers focus on infrastructure (containers, VMs, networking).

This project focuses specifically on coding agent orchestration: session management, HITL (human-in-the-loop) flows, and universal event schemas. These concerns are complementary.

Does it support [platform]? The server is a single Rust binary that runs anywhere with a curl install. If your platform can run Linux binaries (Docker, VMs, etc.), it works. See the deployment guides for E2B, Daytona, Vercel Sandboxes, and Docker.

Can I use this with my personal API keys? Yes. Use sandbox-agent credentials extract-env to extract API keys from your local agent configs (Claude Code, Codex, OpenCode, Amp) and pass them to the sandbox environment.

Why Rust? TODO

Why not use stdio/JSON-RPC?

• has benefit of not having to listen on a port
• more difficult to interact with, harder to analyze, doesn't support inspector for debugging
• may add at some point
• codex does this. claude sort of does this.

Why not OpenCode?

• the harnesses do a lot of heavy lifting
• the difference between opencode, claude, and codex is vast & vastly opinionated