Agent frameworks & infrastructure
Agent sandboxes & runtimes
1,721 agents · most starred first
Agent sandboxes & runtimes, page 8 of 18
The fleet and facilities console for agent runtimes: attach to runtimes you already run, or provision new ones, across machines and harnesses.
HakoVM: Firecracker-class microVMs for macOS, built for AI agents. One binary on Apple's Containerization framework.
Pocket-sized AI agent for embedded Linux in pure C — OpenAI-compatible, tool calling, streaming, ~550KB binary. Runs on routers (MT762X/OpenWrt verified).
🦞 Industrial-grade Soft Execution Engine for OpenClaw. Features Multi-Expert Routing (DeepSeek-R1/V3), DAG Scheduling, API Key Pooling, and Docker Sandbox isolation.
E2BGateway is a unified gateway that enables the official E2B client to interoperate with multiple agent runtimes including agent-sandbox and OpenSandbox, delivering standardized sandbox APIs for AI agents.
The Linux of AI - an open, federated, self-hosted AI operating system. Humans, AI agents, and local LLMs share apps, knowledge, and capabilities on nodes you own. MIT.
AegisIDE: capability-secured agent development environment for local AI models. Rust.
The agent harness for work that cannot leave the building. On-prem, air-gap capable; one static Go binary; sandboxed, policy-governed, every action recorded.
Genus OS — self-hosted AI agent operating system: deterministic engine, plugin-style skills and connectors, multi-tenant CRM + memory. Everything is a plugin.
Run short-lived Python, Node, and Bash with hard limits, live output, and verifiable receipts.
Durable execution and guardrails for AI agents
Sandboxed macOS VMs with full Xcode for AI coding agents — plus a lightweight Linux container mode.
Disposable Linux VMs for your own code. Runs microVMs on Linux, macOS and Windows hosts with no Docker or daemon, GPU through Vulkan, and host folders mounted in. Python, Node.js and Rust.
Self-hosted production platform for deploying, routing, scheduling, and observing Eve agent fleets.
Linura — The intelligent system layer for Linux, AI agents, containers, and VMs.
Kubernetes runtime for multi-user DeepSeek Harness: isolated user environments with persistent storage, a Go launcher and a private platform Connector.
A web console for OpenShell. Everything you do with the openshell CLI (sandboxes, policies, proposals, providers, workspaces) from your browser, without giving up its security model.
Paperclip clone with focus and security and resource consumption in Rust
AI Sandbox
🦞 Dockerized Clawdboss — run a hardened, multi-agent OpenClaw setup in a container. Zero host dependencies.
Docker-powered Ubuntu desktop with OpenClaw pre-installed — access via browser (NoVNC), RDP, or VNC. AI agent gateway ready out of the box.
Desktop sandbox for AI agents and MCP servers — with Zero Token Architecture so your API keys never touch the agent.
Naia OS - Next AI Networks OS
Open-source Agent OS in Rust for long-running autonomous workflows: memory, tool sandboxing, and multi-provider LLM routing.
A modular AI agent built on WASM components. Every capability is sandboxed and swappable.
Atoll is a self-hosted managment plane for AI helpers with a reusable identity catalog, and runtime orchestration.
trama – Agents don't need frameworks. They need a runtime.
A capability-based operating layer for long-running autonomous software engineering agents. Rust daemon, signed capabilities, append-only audit, drift-aware memory, fail-closed sandbox dispatch, commit-scoped provenance.
🌪️ Agent-native orchestration. Declarative, self-hostable runtime for AI agents on any Linux host with containerd.
OmniClaw by Long Leo (Ha Dinh Long) — a Claude Code & Antigravity agent OS governed by 8 immortal Python daemons: zero-trust Git guard, 3-tier MemPalace memory, no agent free-will. Not affiliated with other projects sharing this name.
ESP32-native AI agent runtime with local tools, Lua apps/components, OTA, and web/UART/Telegram control surfaces.
An opensource agentic system infrastructure - not an OS
Self-contained multi-agent AI orchestration runtime for .NET. CLI + web UI, native tools, sandboxing, MCP, layered memory.
the universal realtime, efficient agentic operating system to power future devices like smart glasses
Boot a real x86 Alpine Linux VM in your browser (v86), attach xterm consoles, and let a local AI agent run tools inside the guest—Transformers.js/Ollama, static SPA, no cloud VM required.
Self-hosted runtime for personal apps that combine reliable code with LLM agents and adapt across every screen.
在網路受限的 Docker 容器(Dev Container)中安全執行 Claude Code 的沙箱範本:透過 iptables/ipset 防火牆白名單限制對外連線,隔離 AI Agent 的讀寫與執行。改自 Anthropic 官方 .devcontainer。
A local-first launcher for discovering, installing, running, and managing containerized AI agents and harnesses.
A statically-typed scripting language and bytecode VM for sandboxed execution of AI-generated code, built in C++ with no GC or JIT.
dokan 導管 — agent-operated deterministic script runtime in Docker. MCP control plane + thin operator UI.
Self-hosted AI workstation for a compatible Linux VPS: agent chat, project workspaces, apps, shared browser, remote desktop, terminal and optional mail. MIT licensed.
Self-hosted platform that runs AI agents like employees: identity, Docker sandbox, brokered credentials and a backlog per agent, with guard rails outside the model. Go + Postgres, AGPL-3.0.
A persistent, browser-accessible Linux computer for a single Buzz agent — its own desktop, terminal, filesystem, and long-lived state, joining any Buzz workspace over its relay
在手机上独立运行完整 DeepSeek Harness(DSH)的开源实现。功能已验证,UI 仍在优化中。
HiveGate AI agent runtime — run your Agno agents as a production REST API: per-user sessions and OAuth, human approval, multi-agent teams. (formerly Agents Gateway)
Grok Bot OS: an open-source OS for Super Intelligence, where Grok Bot controls the whole machine
A neutral contract and Go reference implementation for durable agent sessions: replay a run byte-for-byte with no model call, fork it for free, and verify the record independently.
QEMU harness for AI agents: boot aarch64/arm/riscv64/x86_64 kernels in seconds, exec via guest agent, fast rootfs rebuilds, IoT emulation
Multi-tenant AI agents on AWS Lambda MicroVMs — one Firecracker VM per tenant with near-zero idle cost, snapshot resume, auto lifecycle, and EFS-persisted state.
ParseOS is an open-source initiative to advance automation at the OS level. We do not pay for research; instead, our focus is on creating value by helping students and professionals start strong in data science and machine learning. Our goal is to foster a positive shift in tech through collaboration and shared learning.
Isolated development environments for AI agents
Single, consistent interface to run code, manage files, and control isolated environments across sandbox providers
Runs AI coding agents inside isolated Docker containers
Open-source, self-hostable runtime for managed AI agents. Apache-2.0.
A secure runtime for self-correcting AI agents with Docker sandboxing.
Benchmarked agent execution runtime for Python. Sub-10ms cold starts, real-time streaming, time-travel debugging, and self-growing tool libraries. Compare 3 sandbox backends: Docker (OpenSandbox), MicroVM, and in-process AST.
🚀 Веб-интерфейс для AI-агента с возможностью выполнения Python-кода, управления файлами и работы с веб-контентом
Secure runtime for multi-agent AI. Kernel sandboxing (seccomp-bpf), real-time PII redaction, Merkle audit trails.
Sandbox npm, pip, and AI coding agents in ephemeral macOS micro-VMs
KAIROS-ARK is a high-performance, Rust-based Agent Runtime Kernel built for industrial-grade reliability. It delivers sub-100µs dispatch latency, event-sourced deterministic replay, and kernel-enforced capability sandboxing, bridging Python prototypes and production AI systems.
AI agents powered by Claude, running in Docker containers. Desktop app, Slack, and Discord — with chat, terminal, file editor, and custom layouts.
Java-native runtime for long-running AI agents with live execution control, human approvals, and audit trails.
Core Rust library for Lens sandbox policy enforcement, networking, DNS, proxying, and boundary credential exchange.
Claude-In-A-Box: Manage coding agent instances (Claude Code, Codex, Gemini CLI, Cursor) as local processes or containers. REST API, CLI, desktop app.
Open-source managed-agent platform with Docker Compose local runtime/sandbox pools, sessions, vaults, SDK, CLI, and provider adapters.
Graph harness for AI agents: a governed runtime on bipartite graphs (places and transitions) that executes models with scoped permissions, durable state and replayable history.
The next-generation Agent Infrastructure Cybernetics Framework (AICF). Gateway controls, model deployment, app/harness designing and runtime infrastructure management included. It's all on your device!
Standalone Ocean runtime history. Shared development has moved to KINGMAKER-SYSTEMS/ocean.
Portable Rust Agent Client Protocol (ACP) server with model routing, tools, permissions, sandboxing, sessions, and MCP integration.
Build desktop apps with Bun
Long-horizon agent execution harness — reliable autonomous runs for Claude Code, Codex, OpenHands, and custom agents. Goal graphs, spin detection, HITL gates, fork/merge, 8 strategies, 6 validators.
A governed user-space runtime and control plane for long-lived AI agents, with lifecycle, scheduling, durable context, tool authorization, quotas and operator interfaces.
Official SDKs, API clients, and CLI/MCP tools for Daytona, the secure and elastic infrastructure for running AI-generated code
JupyterHub-style multi-user front for DeepSeek Harness (dsh): PAM login, per-user isolated instances, cookie-routed HTTP/WS proxy — zero modifications to dsh.
The incremental build system for AI applications.
GenPark AI Agent Skill - In-memory runtime hotpatching and method swapper allowing dynamic agent self-adaptation with deterministic rollback checkpoints.
Autonomous Agent Lock-Free Treiber Stack with Atomic Compare-and-Swap (CAS) & ABA Mitigation
Pong is local-first version control and execution infrastructure for AI agents. It complements Git by making agent workspaces, state, and handoffs traceable, recoverable, and safe to resume.
E2B-compatible node service and multi-node control plane for Kuasar Sandbox.
Agent-native OS with capability-based security, semantic IPC, and first-class GPU. Offline-first. Preview on Windows/Linux (NVIDIA).
Isolation for AI agents like Claude Code or GitHub Copilot on native Windows to use native buildchain, no docker/WSL.
融合版 Android DSH:预置 bionic Node 24 + PRoot 双运行时、22 个 android_* 设备工具、离线语音唤醒与桌宠。独立包名 io.dsh.fusion,可与其它 DSH 安卓版并排安装。
A Rust platform for AI agents to share and execute tasks.
A runtime where LLM inference is the primitive unit of computation — conversations are processes, prompts are in-memory data, a system image is one JSON | 以 LLM 推理为原始计算单元的运行时
KREASYS: A professional, autonomous, browser-native AI ecosystem and IDE. It features local LLM inference via WebGPU, a persistent Virtual File System (VFS), and multi-modal interaction (text, image, audio, video). Operates entirely client-side for maximum privacy and performance with integrated Telegram delegation.
A System 2 Cognitive Middleware that replaces static data pipelines with runtime code generation. Features Gemini 3 Pro, E2B Sandboxing, and Self-Correcting Triangulation.
Isolated per-project sandboxes for agentic LLMs
A local runtime for AI agents.
Secure Docker wrapper for AI coding agents with filesystem, privilege, and network isolation
An agent-native operating system for the AI era. OpenOS aims to build an open AOS (Agent Operating System) that provides a complete closed loop — from definition to runtime — for long-running agent workloads, supporting the entire lifecycle from development and delivery to production operations.
Local Sandbox Runtime for AI Agents. Run AI-generated code safely with OS, Wasm, and microVM isolation.
An interactive LLM NPC sandbox for everyone using local Ollama or LM Studio agents!
NCZ Distribution Core
Local task engine. Each task leaves a file you can open.
Confine an AI coding agent (oh-my-pi/omp) to one directory with Landlock, and watch it live over eBPF. Experimental.
A platform for running sandboxed containers
Pi Arcweld, an auditable local workspace for Pi: pinned upstream source, curated extensions, bounded MCP tooling, and a reproducible external runtime (pi-arcweld)
Docker-based, multi-tenant platform for autonomous AI agent ecosystems — runs on local or cloud models, with built-in cross-instance federation. Source-available under the PolyForm Small Business License.
RAT: RunAnyThing via Fully Automated Environment Configuration (EMNLP 2026)
The open-source, self-hosted backend for agent-built, agent-powered SaaS—auth, Postgres, AI runtime, sandbox, browser, jobs, billing, security, and observability in one stack.