Sandboxes and secure execution tools for agent evaluation
Isolated sandboxes where agents can run code and tools away from your machine and data. Of the 426 tools on our evaluating agents page, 9 fit this narrower job. By pricing: 6 open source, 2 free or freemium, 1 pricing on the maker's site. The most-starred open repos here are Cua (31,341 stars), LangChain Deep Agents (30,785 stars), rLLM (5,925 stars). Most come from the frameworks (3), evals & observability (3), browser/computer use (1) categories.
Updated
Cua
Open-source infrastructure for computer-use agents: sandboxes, drivers across operating systems and benchmarks.
Added
LangChain Deep Agents
Deep Agents is LangChain's MIT-licensed, batteries-included agent harness for long-horizon, multi-step work: an opinionated agent that runs out of the box and lets you override or replace any piece. It provides sub-agents with isolated context windows, a pluggable filesystem (local, sandboxed or remote), context summarisation and offloading, shell access in your sandbox, persistent memory and human-in-the-loop approvals. It is model-agnostic for any tool-calling LLM and builds on LangGraph, with tracing, evals and deployment through LangSmith.
Added
rLLM
rLLM is an Apache-2.0 framework for training language agents with reinforcement learning. You bring any harness, including 10+ CLI agents such as Claude Code, Codex, mini-swe-agent and OpenCode, or wrap your own LangGraph or OpenAI Agents SDK agent with @rllm.rollout, run it in Docker, Daytona, Modal or local sandboxes, and switch training backends with one flag. The same agent code drives both evaluation and training.
Added
ToolHive
Stacklok platform that runs MCP servers in isolated containers with per-request identity, access policy and observability.
Added
Sage
Sage is an open-source (MIT) RSI agent harness, where RSI stands for Recursive Self-Improvement: it improves agents using feedback from their tasks. Starting from a goal or a curated usage history, it runs the agent, has a separate evaluator grade the results, tunes the agent and keeps every version with its evidence, so you decide which one to adopt. It comes as a Desktop app, a multi-user Server with Agent Studio, and the SAgents Python runtime, with support for Skills, MCP, multi-agent setups and sandboxed runs with approvals.
Added
AgentEnv Framework
AgentEnv Framework is Scale AI's open-source Python SDK and CLI for building RL environments for agents: containerized worlds with realistic apps and data behind an MCP gateway, plus tasks (as DAGs) that grade any agent in any sandbox.
Added
Runtm
Open-source sandboxes where coding agents build and deploy. Spin up isolated environments where Claude Code, Cursor, and other agents code and deploy software.
Added
CoreWeave Forge
One product to run, observe, evaluate and post-train agents (W&B, OpenPipe, sandboxes).
Added
Solari
Solari gives AI agents three kinds of cloud environment from one API: headless cloud browsers that start in about 8 ms for browsing, logging in and scraping, isolated sandboxes that start in about 90 ms for running code and tools, and full computer desktops an agent can drive like a person. Sessions are billed per second while they run, with replay recordings and stealth mode on paid plans. Solari publishes its own latency benchmark against Browserbase, Steel, Kernel and Hyperbrowser. A free plan covers prototyping; Starter is $20 a month and Professional $200.
Added
GitHub repos (7)
| Repo | GitHub stars | Licence | Last push |
|---|---|---|---|
| 31,341 | MIT | 2026-10-30 | |
| 30,785 | MIT | 2026-10-30 | |
| 5,925 | — | — | |
| 2,316 | Apache 2.0 | 2026-10-30 | |
| 1,216 | — | 2026-10-30 | |
| 394 | — | 2026-10-30 | |
| 304 | — | — |
7 repos. Stars and licence come from each repo, checked when the listing was last verified.
FAQ
What are the best AI tools for evaluating agents focused on sandboxes and secure execution?
We list 9. The first three by GitHub stars and featured status are Cua: Open-source infrastructure for computer-use agents: sandboxes, drivers across operating systems and benchmarks.; LangChain Deep Agents: Deep Agents is LangChain's MIT-licensed, batteries-included agent harness for long-horizon, multi-step work: an opinionated agent that runs out of the box and lets you override or replace any piece. It provides sub-agents with isolated context windows, a pluggable filesystem (local, sandboxed or remote), context summarisation and offloading, shell access in your sandbox, persistent memory and human-in-the-loop approvals. It is model-agnostic for any tool-calling LLM and builds on LangGraph, with tracing, evals and deployment through LangSmith.; rLLM: rLLM is an Apache-2.0 framework for training language agents with reinforcement learning. You bring any harness, including 10+ CLI agents such as Claude Code, Codex, mini-swe-agent and OpenCode, or wrap your own LangGraph or OpenAI Agents SDK agent with @rllm.rollout, run it in Docker, Daytona, Modal or local sandboxes, and switch training backends with one flag. The same agent code drives both evaluation and training.
Are there free or open-source options?
Yes: 6 open source, 2 free or freemium, 1 pricing on the maker's site. Pricing is as each maker states it.
More from evaluating agents
Tools are grouped by what each says it does. Missing one? Submit it free.