Sandbox Agent skill for Claude Code: set up and use Rivet's sandboxed coding-agent API

Rivet's skill that teaches Claude to set up Sandbox Agent and drive coding agents like Claude Code and Codex inside sandboxes.

Installs: 10.6K; on skills.sh
Stars: 45; on GitHub
Files: 60+; besides SKILL.md
Updated: Oct 2026; last repo commit

Sandbox Agent is a server with one API for running coding agents (Claude Code, Codex, OpenCode, Amp) in sandboxes. This skill gives Claude the docs for it. Claude can then help you install and start the server, create sessions with the TypeScript SDK or HTTP API, send prompts and read the event stream.

The skill is a SKILL.md plus 216 reference files, all Markdown. It bundles no scripts, but it walks Claude through running npx, curl | sh, Docker and cargo commands, and the server needs the API keys of the agents you run. It also covers deploying to E2B, Daytona, Vercel, Cloudflare and other providers, and the server sends anonymous telemetry by default.

Good for

Developers building a product that runs coding agents in E2B, Daytona or Docker sandboxes
Teams building a chat UI for coding agents on top of one API
Anyone who needs sessions, event streaming and human-in-the-loop prompts for agents

Install sandbox-agent

Claude Code

Copy this prompt.

Install the sandbox-agent skill in Claude Code. Run this command:

npx skills add rivet-dev/skills -s sandbox-agent

Then check that it loaded and tell me how to use it.

2
Paste it into Claude Code. It shows the change and asks before running it.

Requires Node.js.

Installing the skill only gives Claude the Sandbox Agent docs. The server it describes is a separate program, so also run the server with the second method. SKILL.md pins the 0.5.x release line. --no-token --host 0.0.0.0 lets any network interface connect without a token; SKILL.md says that is for sandboxes where the provider secures the network, and to use --token if you expose the server publicly. Our advice: on your own laptop use --host 127.0.0.1 instead of 0.0.0.0. The server sends a small anonymous telemetry payload unless you add --no-telemetry.

From the vendor's docs. Source.

What's in the skill

When it loads

Deploy, configure, and integrate Sandbox Agent - a universal API for orchestrating AI coding agents (Claude Code, Codex, OpenCode, Amp) in sandboxed environments. Use when setting up sandbox-agent server locally or in cloud sandboxes (E2B, Daytona, Docker), creating and managing agent sessions via SDK or API, streaming agent events and handling human-in-the-loop interactions, building chat UIs for coding agents, or understanding the universal schema for agent responses.

Sections in SKILL.md

What To Do If You Get Stuck
Quickstart
Install skill (optional)
Set environment variables
Run the server
Install agents (optional)
Install desktop dependencies (optional, Linux only)
Create a session
Send a message
Read events
Test with Inspector
Next steps
Reference Map
Agents
AI
Api Reference

Key facts

Author: Rivet (rivet-dev). The repo README says this skill is generated and published by rivet-dev/agents, so this is the vendor's own repo.
License: No LICENSE file in the rivet-dev/skills repo; GitHub shows no license and the skill sets none
Installs on skills.sh: 10,633 as of Nov 2, 2026
GitHub stars: 45 for the rivet-dev/skills repo, Nov 2, 2026
What is inside: 217 files including SKILL.md: 216 Markdown reference files, no scripts (git tree)
Pinned product version: Sandbox Agent 0.5.x; the Docker example uses rivetdev/sandbox-agent:0.5.4-full
Last commit to the repo: Oct 27, 2026
Needs: The Sandbox Agent server (Node.js, Bun, Docker or Cargo) and API keys for the agents you run

What it includes

Tells Claude to send you to Rivet Discord or GitHub issues
Claude hands you a pre-generated report: symptoms, error, what you tried, and environment details such as Sandbox Agent version, SDK, agent and sandbox provider.
Read · What to do if you get stuck
Passes provider API keys to the agents
Shows local shell exports, E2B, Daytona and Docker setups. Mentions extracting keys from local Claude Code or Codex config files with sandbox-agent credentials extract-env --export.
Try: “Set up Sandbox Agent in Docker with my Anthropic key.”
Network · Set environment variables
npx, bunx, curl install script, npm, Bun, Node.js SDK or cargo
Eight ways to start sandbox-agent server on port 2468. Includes curl -fsSL https://releases.rivet.dev/sandbox-agent/0.5.x/install.sh | sh and --no-token --host 0.0.0.0.
Runs code · Run the server
Adds --token auth and points to the CORS guide
SKILL.md says tokens are usually not needed inside provider sandboxes, and to use --token if you expose the server publicly.
Read · Configure a token and CORS
sandbox-agent install-agent --all; install desktop --yes
Agents are otherwise installed on first session. The desktop packages are Linux only and enable the /v1/desktop/* screenshot and input APIs.
Runs code · Install agents and desktop packages
TypeScript SDK createSession, prompt and onEvent
Example code connects to http://127.0.0.1:2468, starts a claude session at cwd /, sends a prompt and prints events.
Runs code · Create a session, send a message, read events
Opens a web UI at /ui/ on the server
Lets you look at sessions and events in a browser.
Read · Test with Inspector
216 reference files in Agents, AI, API reference, Deploy and General
Covers six agents, ten deploy targets (E2B, Daytona, Docker, Vercel, Cloudflare, Modal and more), sessions, persistence, MCP, processes, file system, computer use, security and telemetry.
Read · Reference map

From the official docs

Permissions

High risk

This skill walks Claude through starting a server that can run agents, commands and file changes, and its examples leave that server open on every network interface. Bind it to 127.0.0.1 on your own machine and add a token before you expose it.

What it can do

Read your data
Create and change things
Delete things
Read-only mode

Watch out for

The setup examples start an open server
The examples use --no-token --host 0.0.0.0, so any network interface can reach the server. SKILL.md says to use --token when you expose it publicly. Bind to 127.0.0.1 on your own machine. github.com
It can pipe a remote script into your shell
One install route is curl -fsSL https://releases.rivet.dev/sandbox-agent/0.5.x/install.sh | sh. Prefer the npx or Docker route, or read the script first. github.com
It can pull API keys from your Claude Code config
SKILL.md mentions sandbox-agent credentials extract-env --export, which extracts keys from local Claude Code or Codex config files. Only run it if you want those keys in the sandbox. github.com
Desktop packages install with --yes
SKILL.md's optional desktop step is sandbox-agent install desktop --yes (Linux only), which installs the desktop runtime packages without asking you to confirm. It enables the /v1/desktop/* screenshot, mouse and keyboard APIs. Run it only inside a sandbox you can throw away. github.com
The server sends anonymous telemetry
It sends the version, OS, architecture and sandbox provider at start and every 5 minutes, with a random ID stored on disk. Add --no-telemetry to turn it off. github.com
Agents you start use your provider keys
The agents connect to LLM providers with your keys. SKILL.md points to gateways such as OpenRouter, LiteLLM and Portkey for per-tenant budgets, and to the mock agent for tests. github.com

Security advisories

No published security advisories as of Nov 2, 2026.

Technical details

Turning on read-only	SKILL.md names no read-only setting.
Talks to	releases.rivet.dev (install script). rivet.dev (docs and Discord links). Your LLM providers via the agents
Sign-in and keys	Provider API keys are passed to the agents as environment variables; SKILL.md gives the shell, E2B, Daytona and Docker ways to set them.
Scopes	Walks Claude through running npx, bunx, curl \| sh, Docker and cargo commands The Sandbox Agent API described in the references can run commands, send SIGTERM or SIGKILL to processes, write and delete files, and drive a desktop (screenshots, mouse, keyboard) Passes ANTHROPIC_API_KEY and OPENAI_API_KEY to agents and extracts keys from local Claude Code or Codex config files

Questions

What does the sandbox-agent skill do?

It gives Claude the Sandbox Agent docs, so it can help you run the server, create agent sessions through the SDK or HTTP API, and stream events. Sandbox Agent itself is a separate server that you run.

How do I install it?

Use the install box above. The first method adds the skill only. The server is a separate program, so also use the Run the server methods and set the API keys for the agents you plan to run.

Is it official?

It comes from Rivet's own repo. The repo README says the skill is generated and published by rivet-dev/agents, so edits belong in that repo.

Does it need an API key?

The skill itself names none. SKILL.md says each coding agent needs the key for its provider, such as ANTHROPIC_API_KEY or OPENAI_API_KEY. It also describes a mock agent for tests that need no provider credentials.

Is it safe to follow its setup steps?

The skill bundles no scripts, but the steps it walks Claude through start a server, install agents and can run a remote install script. The examples use --no-token --host 0.0.0.0, which accepts connections on every network interface. Our advice: bind to 127.0.0.1 on your own machine and add a token before you expose it.

Does the server send telemetry?

The bundled telemetry page says it sends a small anonymous payload on startup and every 5 minutes: version, OS name, architecture, OS family and the detected sandbox provider. Each sandbox gets a random ID stored on disk. Add --no-telemetry to turn it off.

Does it work in the Claude app?

This page has no Claude app tab. The skill is about running a local server with Node, Docker or Cargo, which needs a terminal. Use Claude Code or another agent with shell access.

Information

updated Nov 2, 2026

Repository: rivet-dev/skills
Folder: skills/sandbox-agent

rivet-dev/skills skills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 5

sandbox-agent SKILL.md · github.com
rivet-dev/skills README · github.com
skills.sh listing · skills.sh
telemetry.md (bundled reference) · github.com
security.md (bundled reference) · github.com