AIOS: AI Agent Operating System
An operating system that embeds LLMs and manages scheduling, memory, storage, tools, and SDKs for AI agents.
- Source repo
- agiresearch/AIOS
- Stars
- ★ 6.5k
- Last updated
- 3mo ago
- License
- NOASSERTION
- Primary language
- Python
- FA score
- 34/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platform
- Cost
- Free software; you pay for model usage
- Setup effort
- High · needs real infrastructure
- You'll need
- Typical use
- Agent developers who want a unified infrastructure for scheduling, memory, and tool management across multiple agent frameworks.
- Not a fit if
- Users without GPUs or a virtualized GUI environment for computer-use agents
- Developers on Python 3.12+ or non-Linux setups for vLLM
- Source review
- 34/100 · Major gaps 8 safety controls not found
What does this agent do, and when should you use it?
AIOS is an AI Agent Operating System that embeds large language models (LLMs) into the operating system, addressing issues like scheduling, context switching, memory management, storage management, tool management, and Agent SDK management during LLM-based agent development and deployment. It comprises the AIOS Kernel (this repository) and the AIOS SDK (the Cerebrum repository). The kernel abstracts the OS kernel to manage resources such as LLMs, memory, storage, and tools. The SDK supports agent developers and users in building and running agents. AIOS supports multiple deployment modes: Local Kernel Mode, Remote Kernel Mode, and Remote Kernel Dev Mode. It integrates various agent frameworks like ReAct, Reflexion, AutoGen, and MetaGPT, and provides both Web UI and Terminal UI.
The AIOS kernel receives agent queries and processes them through a chain of syscalls, scheduling and dispatching to modules like LLM cores, context manager, memory manager, storage, and tool manager. It supports LLM providers including OpenAI, Anthropic, Gemini, Groq, HuggingFace, ollama, vLLM, Deepseek, and Novita. The terminal UI provides an LLM-based semantic file system for natural language file management. For computer-use agents, the Tool Manager is redesigned as a VM Controller and MCP Server, enabling sandboxed GUI interaction. The kernel can be launched via Uvicorn and offers both Web and terminal interfaces.
- Agent developers who want a unified infrastructure for scheduling, memory, and tool management across multiple agent frameworks.
- Users on resource-constrained devices (e.g., mobile) who need to run or develop agents on a remote high-performance machine using Remote Kernel Mode.
- Developers integrating various agent frameworks (AutoGen, MetaGPT) without rebuilding common components.
- Users who prefer natural language file system commands via the AIOS Terminal instead of traditional shell.
- Developers building computer-use agents that require a sandboxed VM environment to safely execute GUI interactions.
How do you install or deploy this agent?
Clone the AIOS repository and set up a Python 3.10 or 3.11 virtual environment. Install dependencies via uv or pip: for GPU uv pip install -r requirements-cuda.txt, for CPU uv pip install -r requirements.txt. Then clone and install the Cerebrum SDK: git clone https://github.com/agiresearch/Cerebrum.git and cd Cerebrum && uv pip install -e .. Optionally install a virtualized GUI environment (e.g., OSWorld) for computer-use agents.
How do you use this agent?
Edit aios/config/config.yaml to set API keys and LLM models (e.g., OpenAI, Deepseek, ollama). Launch the kernel with bash runtime/launch_kernel.sh or python3.11 -m uvicorn runtime.launch:app --host 0.0.0.0 --port 8000. For the terminal UI, run python scripts/run_terminal.py and type natural language commands. For remote kernel mode, configure Cerebrum with matching host and port. For computer-use agents, refer to the Cerebrum documentation.
What are this agent's strengths and limitations?
- Unified platform for LLM scheduling, memory, storage, and tool management, saving developer effort.
- Support for multiple LLM backends (OpenAI, Anthropic, HuggingFace, vLLM, etc.) reducing vendor lock-in.
- Integration with multiple agent frameworks (AutoGen, MetaGPT) out of the box.
- Remote kernel mode enables use on low-resource devices.
- Optional Rust rewrite (aios-rs) for performance-focused components.
- Installation and configuration are complex, requiring multiple steps and API keys.
- Only supports Python 3.10 and 3.11, restricting some environments.
- Currently only Mode 1 and Mode 2 are fully supported; other modes are in development.
- Computer-use agents require an additional virtualized GUI environment.
- Documentation may not cover all features, and users may need to explore the source.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| AIOS: AI Agent Operating System This agent | 34 · Major gaps | Self-hosted serviceFree + model costs | ★ 6.5k | 3mo ago | Python | — |
| PowerContext | 55 · Major gaps | CLIFree | ★ 1.3k | today | Python | Codex · Claude Code |
| Beads: Branch-Aware Memory for Coding Agents | 53 · Major gaps | CLIFree | ★ 28k | today | Go | Codex · Claude Code |
| OSWorld 2.1 Computer-Use Agent Benchmark | 53 · Major gaps | CLIFree + model costs | ★ 376 | 4d ago | Python | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows: README instructs users to write API keys directly into config file, no mention of least privilege or permission isolation; no user confirmation mechanism; data flow is opaque, no explanation of how data is processed or transmitted; sensitive data handling lacks description; dependencies are not pinned, supply chain risk; external effects not explicit but network calls and file operations exist; rollback feature only mentioned as requiring redis, no mechanism details; source attribution only via papers and repo links, publisher identity unverified. Therefore all trust criteria scored 0.
Evidence shows: README and code structure are consistent, module division clear, good self-consistency; dependency list includes many unpinned packages and relies on external services (e.g., redis, ollama), availability risk; failure messages not explicit in docs, but test files have error handling. Thus self-consistency 2, dependency availability 1, failure messages 1.
Evidence shows: README identifies target users (developers, users) and multiple deployment modes, broad scenario coverage; capability boundaries via module division and docs, but no explicit limits; trigger precision not detailed, only natural language commands; environment fit provides CPU/GPU install options, but not all platform limitations. Thus scores 2,2,1,2 respectively.
Evidence shows: README structure clear, includes architecture, installation, quickstart; install notes detailed but rely on external repos; naming stability not explicit but version numbers exist; examples and FAQ have some examples but no FAQ; known limitations not explicitly listed; license file empty, cannot confirm; versioning changelog not provided; maintenance responsibility not explicit. Thus scores 2,2,1,2,1,0,1,1 respectively.
Evidence shows: output usability not explicit but terminal and Web UI provided; marginal value high, multiple features; cost-benefit not assessed, but requires multiple API keys and GPU resources. Thus scores 1,2,1 respectively.
Evidence shows: README cites multiple papers but no implementation details; cross-source verification insufficient, relies only on repo itself; fact and inference not clearly separated. Thus scores 2,1,1 respectively.
- Not found in source: least-privilege scopingGrant only what the task needs: a dedicated account or read-only token, scoped to specific directories and repos.
- Not found in source: confirmation before actingTurn on (or add) a confirmation step before it acts, and try it in a sandbox or test environment before real data.
- Not found in source: data-flow disclosureWatch which external services it contacts (proxy or firewall logs) and keep sensitive data out until you know where it goes.
- Not found in source: sensitive-data handlingUse dedicated, low-privilege, revocable API keys — never production credentials — and keep secrets out of logs.
- Not found in source: dependency securityPin versions and run a dependency audit (npm audit, pip-audit) before installing; prefer running it in a container.
- Not found in source: disclosed external effectsEstablish which external systems it writes to, sends to or changes, and verify with test accounts or repos before production.
- Not found in source: rollback or recovery pathBack up first, or work on a git branch or snapshot, so its changes can be undone.
- Not found in source: verifiable attributionInstall from the official repo or registry and check the publisher and URL to avoid look-alike packages.
- API keys are written directly into config file, risk of leakage; recommend using environment variables or secret management.
- Dependencies are not pinned, supply chain attack risk; recommend locking versions and regular updates.
- License file is empty, open-source license cannot be confirmed; verify before use.
- Rollback feature depends on redis, data persistence mechanism not explained, may lead to data loss.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
What's Changed * Updated the comment in llm_core by @Luckfort in https://github.com/agiresearch/AIOS/pull/465 * update apis for launching LSFS terminal by @dongyuanjushi in https://github.com/agiresearch/AIOS/pull/466 * add patch for fixing LSFS apis by @dongyuanjushi in https://github.com/agiresearch/AIOS/pull/467 * initial commit by @WujiangXu in https://github.com/agiresearch/AIOS/pull/468 * update temperature and max new tokens by @dongyuanjushi in https://github.com/agiresearch/AIOS/pull/469 …View on GitHub ↗ -
Release notes excerpt
What's Changed * feat:Improve API Key Error Handling by @XiangZhang-zx in https://github.com/agiresearch/AIOS/pull/428 * docs: Update installation steps for AIOS Kernel and SDK by @XiangZhang-zx in https://github.com/agiresearch/AIOS/pull/429 * add deepseek support by @dongyuanjushi in https://github.com/agiresearch/AIOS/pull/431 * update deepseek docs by @dongyuanjushi in https://github.com/agiresearch/AIOS/pull/433 * feat: integrate semantic filesystem by @dongyuanjushi in https://github.com/agiresearch/AIOS/pull/436 …View on GitHub ↗ -
Release notes excerpt
What's Changed * Add exception messages by @dongyuanjushi in https://github.com/agiresearch/AIOS/pull/370 * add cors for launching kernel by @dongyuanjushi in https://github.com/agiresearch/AIOS/pull/371 * feat: Enhanced Error Logging by @XiangZhang-zx in https://github.com/agiresearch/AIOS/pull/372 * docs: update README for v0.2.0.beta release by @yongfeng101 in https://github.com/agiresearch/AIOS/pull/373 * update README.md by @dongyuanjushi in https://github.com/agiresearch/AIOS/pull/375 …View on GitHub ↗
FAQ
What is the relationship between AIOS and Cerebrum?
Which API keys are required to deploy AIOS?
config.yaml.Does AIOS support agent frameworks?
Can I use AIOS without a GPU?
requirements.txt. You can also use remote kernel mode to access a GPU server.