The 10 best open-source AI agents you can run yourself in 2026
An agent does not just answer — it takes steps: reads files, runs commands, calls other systems. These are the open-source ones worth knowing, grouped by the job they do.
Short answer
Choose by job. For a do-anything personal assistant: OpenClaw. For clerical office work with exact calculations: OpenClerq. For questions across your documents: AnythingLLM. For coding: OpenHands, Goose, Cline or Aider. For workflows between business apps: n8n (fair-code) or Activepieces. For letting a model run code on your computer: Open Interpreter. All of them run on hardware you control, with the model you choose.
We build OpenClerq, which appears on this page. We have tried to describe every tool fairly and to say where another one is the better choice. Facts checked 19 September 2026; if something is wrong, open an issue.
The list at a glance
| Agent | Job | Licence | Interface |
|---|---|---|---|
| OpenClaw | Personal assistant for almost anything | MIT | Messaging apps |
| OpenClerq | Clerical work with exact calculations | MIT | Desktop app, local API |
| AnythingLLM | Questions across your documents | MIT | Desktop app, or a server |
| OpenHands | Software development tasks | MIT | Web interface, command line |
| Goose | Coding and general tasks through extensions | Apache 2.0 | Desktop app, command line |
| Cline | Coding inside VS Code, step by step | Apache 2.0 | Editor extension |
| Aider | Pair programming in a git repository | Apache 2.0 | Command line |
| n8n | AI workflows between business apps | Sustainable Use License (fair-code) | Browser workflow editor |
| Activepieces | Zapier-style flows with AI steps | MIT (Community Edition) | Browser flow builder |
| Open Interpreter | Letting a model run code on your computer | AGPL-3.0 | Command line |
How we chose: the source is public, you can run it on your own hardware, it works with more than one model provider including local models, and it was actively maintained when we checked. n8n is source-available rather than OSI open source; it is here because it is the most widely self-hosted agent workflow tool.
Personal and office agents
01OpenClaw
- Best for
- One assistant for email, calendar, files, scripts and the web
- Licence
- MIT
The breakout open-source agent of 2026: it runs on your own machine and you talk to it through WhatsApp, Telegram, Signal or Discord. It can read and write files, run scripts and control a browser, and a large community publishes skills for it.
Watch out: that reach is also its risk. Researchers have found third-party skills that exfiltrated data. Isolate it and review what you install. OpenClerq vs OpenClaw.
02OpenClerq
- Best for
- Recurring clerical work on sensitive data
- Licence
- MIT
Our product. A desktop agent for office work with a narrow, authenticated gateway on your own machine. Calculations run in a deterministic engine; the model — local or one of six cloud providers — explains. Schedules, watched files and webhooks run tasks on their own, and every run is recorded with its cost.
Watch out: at release 0.5 its model does not yet call tools on its own; the sandboxed tool loop arrives in 0.6.
03AnythingLLM
- Best for
- Agents that work across your documents
- Licence
- MIT
Workspaces over your own files, with agents that can search, summarise and use tools. Runs as a single-user desktop app or as a multi-user server, with local or cloud models.
Watch out: agent features depend heavily on the model you connect; small local models struggle with multi-step tool use.
Coding agents
04OpenHands
- Best for
- Handing over whole development tasks
- Licence
- MIT
Formerly OpenDevin. Agents that write code, run commands and browse, working inside a sandboxed container. Use it from a local web interface or the command line, or through its hosted service.
05Goose
- Best for
- A general agent extended through MCP servers
- Licence
- Apache 2.0
Created by Block and contributed to the Linux Foundation's Agentic AI Foundation. Runs locally as a desktop app or on the command line, works with many model providers, and gains abilities through Model Context Protocol extensions.
06Cline
- Best for
- Supervised coding inside the editor
- Licence
- Apache 2.0
An agent in VS Code that plans, edits files and runs commands, asking for approval at each step. Bring any model, including local ones.
07Aider
- Best for
- Pair programming from the terminal
- Licence
- Apache 2.0
Edits code in your git repository and commits each change with a sensible message, so every step can be reviewed or undone. Works with most cloud and local models.
Workflow agents
08n8n
- Best for
- AI steps inside workflows between business apps
- Licence
- Sustainable Use License — free to self-host for internal use
A visual workflow platform with hundreds of integrations and AI agent, chain and vector store nodes. Self-host it or use n8n's cloud. OpenClerq vs n8n.
09Activepieces
- Best for
- An MIT-licensed, Zapier-style builder with AI agents
- Licence
- MIT for the Community Edition; enterprise features are commercial
A no-code flow builder with AI agents and support for MCP servers, self-hostable with Docker. The closest permissively licensed equivalent to n8n.
Computer-control agents
10Open Interpreter
- Best for
- Letting a model run Python, JavaScript and shell commands for you
- Licence
- AGPL-3.0
A chat in your terminal where the model writes code and runs it on your computer — to convert files, analyse data or automate the desktop. It asks before running code by default; keep it that way.
Before you let any agent act
- Sandbox it. A container, a separate user account, or at least a folder it cannot leave.
- Grant the least it needs. Specific folders, specific hosts, specific commands.
- Require approval for anything irreversible — sending, deleting, paying, publishing.
- Set a budget and a timeout. No loop should be able to run all night or spend without limit.
- Keep a log of every step and model call, and read it.
- Assume prompt injection. Anything an agent reads — a web page, an email, a document — can contain instructions aimed at it.
Frequently asked questions
What is the difference between an AI agent and a chatbot?
A chatbot answers. An agent works in a loop: it decides on a step, uses a tool — reading a file, running a command, calling an API — looks at the result, and decides the next step until the task is done.
Are open-source AI agents safe?
As safe as the reach you give them. Open source lets you read exactly what an agent can do, but an agent with shell access and an inbound channel is still a large attack surface. Sandbox it, limit its permissions and keep approvals on.
Which open-source agent is best for business tasks?
For connecting business apps, n8n or Activepieces. For careful clerical work on sensitive data, OpenClerq. For a general assistant, OpenClaw — with care.
Can open-source agents use local models?
All ten can, usually through Ollama or any OpenAI-compatible server. Multi-step tool use is demanding, though, and small local models make more mistakes at it than large cloud models.