---
title: "Integrations, agent frameworks and eval harnesses"
description: "Promigence packages for the agent frameworks and eval harnesses you already use, with how to install each one and where to read more."
url: "https://www.promigence.ai/integrations"
site: "Promigence"
---

# Integrations

Promigence packages for the agent frameworks and eval harnesses you already use. Install one, and the work your agent does runs in a Promigence sandbox while the rest of your code stays as it is.

## Agent frameworks

Your agent keeps its framework. Its shell commands, file edits and code run in a Promigence sandbox.

### [LangChain](https://www.promigence.ai/integrations/langchain) (`langchain-promigence`, PyPI)

Use a Promigence sandbox as the backend of a LangChain Deep Agent, so the agent's shell commands and file edits happen in the sandbox. The same package adds a `promigence` sandbox provider to deepagents-code.

### [OpenAI Agents SDK](https://www.promigence.ai/integrations/openai-agents-sdk) (`promigence-openai-agents`, PyPI)

Give an OpenAI Agents SDK `SandboxAgent` a Promigence sandbox, so its shell and file tools run there instead of on your machine. Your agent code stays the same; only the client you pass in changes.

### [Claude Agent SDK](https://www.promigence.ai/integrations/claude-agent-sdk) (`promigence-claude-agent-sdk`, npm)

Keep the Claude Agent SDK loop in your own process and run Claude Code itself inside a Promigence sandbox. Anything the agent edits, installs or deletes stays in the sandbox.

### [Vercel AI SDK](https://www.promigence.ai/integrations/vercel-ai-sdk) (`promigence-ai-sdk-sandbox`, npm)

Pass a Promigence sandbox to the Vercel AI SDK as `experimental_sandbox`, or to bash-tool. The commands your tools run and the files they touch stay in the sandbox, away from your server.

### [Google ADK](https://www.promigence.ai/integrations/google-adk) (`promigence-google-adk`, PyPI)

Point a Google ADK agent's environment tools, or its code executor, at a Promigence sandbox. The agent's commands and the code its model writes then run there instead of on the machine serving the agent.

### [Strands Agents](https://www.promigence.ai/integrations/strands) (`strands-promigence`, PyPI)

A Strands agent handed a Promigence sandbox runs its shell and file-editor tools there. Output streams back while a command is still running, so a line a long test run prints early reaches the agent well before the run ends.

### [Claude Managed Agents](https://www.promigence.ai/integrations/claude-managed-agents) (`promigence-claude-managed-agents`, PyPI)

Run the workers of a Claude Managed Agents self-hosted environment on Promigence sandboxes, as a pool you size and give a spend cap. The sessions on that environment are served from the pool.

## Evals and benchmarks

Each sample, trial or task instance gets its own sandbox, and many run at once.

### [Inspect](https://www.promigence.ai/integrations/inspect) (`inspect-promigence`, PyPI)

Use `sandbox="promigence"` in Inspect and every sample gets its own cloud sandbox, with no Docker daemon on your machine. Tasks written for Docker, from a Dockerfile or a single-service compose file, run unchanged.

### [Harbor](https://www.promigence.ai/integrations/harbor) (`promigence-harbor`, PyPI)

Select Promigence in Harbor with one flag. Terminal-Bench and other Harbor tasks then run in cloud sandboxes, each trial in its own copy of the task's environment and many trials at once, with no change to Harbor.

### [mini-swe-agent](https://www.promigence.ai/integrations/mini-swe-agent) (`promigence-mini-swe-agent`, PyPI)

Run mini-swe-agent's SWE-bench instances on Promigence sandboxes. Nothing in mini-swe-agent changes: you name the Promigence environment class on the command line, and each instance runs its commands in its own sandbox while many workers run at once.

Building your own harness? Start sandboxes from the [Python SDK](https://www.promigence.ai/docs/sdk-python), the [TypeScript SDK](https://www.promigence.ai/docs/sdk) or the [REST API](https://www.promigence.ai/docs/reference/api). The [guides](https://www.promigence.ai/docs/guides) go step by step.