Best Open-Source AI Agents in 2026: 11 Picks Compared

The best open-source AI agents in 2026 are Hermes Agent and Goose for a general agent you run yourself, Open Interpreter and DeepSeek Harness for coding and terminal work, DeerFlow for long research jobs, Browser Use, Nanobrowser and MagenticLite for browser tasks, and LangGraph, CrewAI and the OpenAI Agents SDK if you would rather build your own agent.

Which one is best depends on the job, so this guide sorts them by type. Every pick has an open-source licence, a recent commit and an AgenticMart listing that shows what it can touch. Install commands are quoted from each project's README and were not run for this post.

The picks at a glance

Stars as of 12 November 2026, from each repository's GitHub page. Stars show attention, not quality.

Agent	Type	Stars	Licence	Where it runs	Listing
Hermes Agent	General	262,916	MIT	Your machine, a VPS, Docker, SSH or serverless backends	/agents/hermes-agent
Goose	General and coding	56,498	Apache-2.0	Your machine (desktop app, CLI, API)	/agents/goose
Open Interpreter	Coding	68,874	Apache-2.0	Your terminal	/agents/openinterpreter
DeepSeek Harness	General harness (preview)	279,684	MIT	Your machine, local web UI	/agents/deepseek-harness
DeerFlow	Research	85,285	MIT	Your machine or server, Docker recommended	/agents/deer-flow
Browser Use	Browser	120,836	MIT	Local library or CLI, optional paid cloud	/agents/browser-use
Nanobrowser	Browser	14,292	Apache-2.0	Chrome or Edge extension	/agents/nanobrowser
MagenticLite (Magentic-UI)	Browser and files	10,114	MIT	Local web app on macOS or Windows (WSL)	/agents/magentic-ui
LangGraph	Framework	44,864	MIT	Library in your code	/agents/langgraph
CrewAI	Framework	60,816	MIT	Library and CLI, optional paid control plane	/agents/crewai
OpenAI Agents SDK	Framework	30,639	MIT	Library in your code	/agents/openai-agents-python

Every one of them needs a model. You bring an API key or point it at a local model server, so "runs locally" means the agent runs on your hardware, not necessarily the model.

How to read "what it can touch"

Each section combines what the project's README says it does with the "What it can touch" panel on its AgenticMart listing, from automated checks on one commit run on 5 November 2026. "Found in the code" means the check saw it. "Not seen" means the check did not find it, and the tool may still do it. For a full pre-install routine, use the is this AI agent safe checklist.

General agents you run yourself

Hermes Agent

Good for: a personal agent that stays running, remembers you across sessions and reaches you in chat apps. The README describes agent-curated memory, skills it creates from experience, a cron scheduler, and one gateway for Telegram, Discord, Slack, WhatsApp, Signal and the CLI. Switch models with hermes model.

How to start (quoted from the docs, not run for this post):

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
source ~/.bashrc
hermes

On Windows the README gives a PowerShell installer instead. Run hermes setup for the full wizard and hermes tools to choose which tools are on.

What it can touch: a terminal on whichever backend you pick. The README lists six: local, Docker, SSH, Modal, Daytona and Vercel Sandbox. It can message you on every platform you connect and can run jobs unattended on a schedule. The listing shows shell commands, dynamic code, environment variables and secrets, and internet access, all found in the code. Pick the Docker or a remote backend if you do not want it on your own disk, and connect only the chat accounts you need.

Goose

Good for: one agent for code and everyday work, as a desktop app, CLI or API. The README says it works with 15+ model providers, connects to 70+ extensions through the Model Context Protocol, and can use an existing model subscription through ACP, the Agent Client Protocol. Goose sits under the Agentic AI Foundation at the Linux Foundation, which matters if you care about who governs a project.

How to start (quoted from the docs, not run for this post): download the desktop app from the installation page, or install the CLI:

curl -fsSL https://github.com/aaif-goose/goose/releases/download/stable/download_cli.sh | bash

What it can touch: whatever its extensions reach, so its reach grows with each MCP server you add. The listing shows shell commands found in the code. File access was not seen by the check, yet an agent that edits code writes files: a good reminder that "not seen" is not "can't". Check what an extension adds in the MCP servers directory before you connect it.

DeepSeek Harness

Good for: developers who want a plugin-first agent harness and can live with breaking changes. The README calls it "everything-is-a-plugin" and says plainly it is in developer preview with compatibility-breaking changes coming. It has the most GitHub stars on this list, which tells you about attention, not maturity.

How to start (quoted from the docs, not run for this post):

npx @deepseek-ai/dsh web

That starts a web UI at http://127.0.0.1:3080 by default.

What it can touch: its own safety notice is unusually direct. It says the project has not had a security audit, can run model-generated code and commands, load third-party plugins, and reach the network, processes, credentials and files available to it. It recommends a disposable VM or container. The listing shows dynamic code, environment and secrets, and internet access found in the code. Take the notice at its word and give it a throwaway environment.

Best AI coding agents, open source

Open Interpreter

Good for: a terminal coding agent tuned for low-cost models. The README says it is a fork of OpenAI's Codex focused on emulating the harness that gets the most out of cheaper models, and /harness switches between styles. It works in ACP-compatible editors through interpreter acp, and it speaks the Codex exec protocol, so the README says a Codex SDK app can swap the binary path to interpreter.

How to start (quoted from the docs, not run for this post):

curl -fsSL https://www.openinterpreter.com/install | sh

Then type i or interpreter to start. Windows uses irm https://www.openinterpreter.com/install.ps1 | iex.

What it can touch: your terminal and the project you run it in. The listing shows dynamic code found in the code. Run it inside the repository you mean to change, not your home folder.

Strong coding agents not on AgenticMart yet

Several of the most-starred open-source coding agents are not listed on AgenticMart as of 12 November 2026, so we cannot show a "what it can touch" panel for them. They are still worth a look. Stars and licences below are from their GitHub pages on 12 November 2026:

Agent	Stars	Licence
OpenCode	226,386	MIT
Codex CLI	135,027	Apache-2.0
OpenHands	93,332	MIT
Cline	71,560	Apache-2.0
Aider	50,089	Apache-2.0

If you maintain one of these, listing it is free.

A 2026 change to know about: Gemini CLI (108,798 stars, Apache-2.0) is still open source, but its README now says it is designed for Gemini Code Assist Standard and Enterprise licences. Free individual access and Google AI Pro and Ultra moved to Antigravity CLI on 18 June 2026. Many lists still recommend Gemini CLI as a free option, which no longer matches its own README.

Research agent

DeerFlow

Good for: long, multi-step research and reports. Version 2.0 is a ground-up rewrite that the README calls a "super agent harness": sub-agents, long-term memory, sandboxed execution, skills and MCP support, built on LangGraph and LangChain. The original research framework lives on the 1.x branch.

How to start (quoted from the docs, not run for this post):

git clone https://github.com/bytedance/deer-flow.git
cd deer-flow
make setup
make docker-init
make docker-start

Then open http://localhost:2026. make setup is an interactive wizard that writes config.yaml and .env.

What it can touch: a file system and sandbox, uploads and outputs, plus any MCP servers you configure. The README tells you to read its security notice before running it on a server or exposing it beyond localhost. The listing shows environment and secrets, and internet access, found in the code. Keep it on localhost unless you have put authentication in front of it.

Browser agents

Browser Use

Good for: developers who want a browser agent in their own code. In version 1.0 the model writes Python against raw Chrome DevTools Protocol in a persistent REPL, according to the README. There are three ways in: a hosted cloud agent (paid), a CLI, and the open-source Python library.

How to start (quoted from the docs, not run for this post), for the library, Python 3.11 or newer:

uv init --python 3.12 my-agent && cd my-agent
uv add 'browser-use==1.0.3'

The README pins a version, which is a habit worth keeping.

What it can touch: a real browser, and if you connect your own, your logged-in sessions. That is the main thing to decide before you start: a fresh browser profile limits the damage if a page tricks the agent. The listing shows browser control (from its description), environment and secrets, and internet access.

Nanobrowser

Good for: people who want a browser agent without writing code. It is a Chrome and Edge extension with a side panel, and you bring your own model key or a local model through Ollama or LM Studio, per the README.

How to start (quoted from the docs, not run for this post): install it from the Chrome Web Store, connect a provider on the welcome page, then press Alt+Shift+N to open the side panel.

What it can touch: the pages you are on and your tabs. Its permissions table is the clearest on this list. In the default Ask mode it reads, clicks and scrolls freely, asks before it submits, sends, deletes or uploads, and hands you the step for payments, passwords and access grants. You can set each site to Allow, Ask or Block. The README also says a daily usage count is sent without an ID. The listing shows browser control from its description.

MagenticLite (Magentic-UI)

Good for: browser and local-file tasks on smaller models, from Microsoft. The README says MagenticLite is the next generation of Magentic-UI, built to pair a small orchestrator model with a browser-use model, and that it checks in before critical actions. The older 0.1 release for frontier models lives on its own branch.

How to start (quoted from the docs, not run for this post):

uv venv --python=3.12 --seed .venv
source .venv/bin/activate
uv pip install "magentic_ui>=0.2.0"
magentic-ui --port 8081

Then open http://127.0.0.1:8081/ and connect a model endpoint.

What it can touch: the web and your local file system. The README says browser sessions run inside a lightweight VM sandbox. The listing shows shell commands, environment and secrets, and internet access found in the code, plus file and browser access from its description.

Open-source AI agent frameworks

Pick a framework when none of the ready-made agents fit and you want control over every step. You write the agent; the framework handles state, tools and orchestration. For agents that need to call each other across systems, read up on A2A.

LangGraph

Good for: long-running, stateful agents where you want to draw the control flow yourself. The README calls it a low-level orchestration framework, and points to Deep Agents as a higher-level package on top of it if you want to move faster.

How to start (quoted from the docs, not run for this post): pip install -U langgraph

What it can touch: only the tools you write or wire in. The listing shows dynamic code found in the code.

CrewAI

Good for: multi-agent teams defined by role. The README splits it into Crews, for autonomous role-based agents, and Flows, for event-driven control. A commercial control plane is sold separately; the framework itself is MIT.

How to start (quoted from the docs, not run for this post):

uv tool install crewai
crewai create crew latest-ai-development
cd latest_ai_development
crewai install
crewai run

What it can touch: the tools you give each agent. CrewAI's listing has not been scanned yet as of 12 November 2026, so it shows no "what it can touch" findings.

OpenAI Agents SDK

Good for: a small, readable framework with handoffs, guardrails, sessions, human-in-the-loop and tracing built in. Despite the name, the README says it is provider-agnostic and supports 100+ other LLMs. Python 3.10 or newer.

How to start (quoted from the docs, not run for this post): pip install openai-agents, then set OPENAI_API_KEY (or configure another provider) before running the examples.

What it can touch: what you give it: function tools, MCP servers and hosted tools. It also offers sandbox agents that work inside a container. The listing shows none of the tracked capabilities seen in the code; for a library that is expected, because your code adds them.

How we picked

These are the criteria. A project had to meet all of them:

1. Open-source licence on GitHub. MIT or Apache-2.0 here. We left out projects whose main product is source-available. For example, the AutoGPT README puts its platform under Polyform Shield, while only its classic code is MIT.
2. An agent or agent framework. No awesome-lists, courses or prompt collections, even popular ones.
3. Actively maintained. We dropped AutoGen: its README says it is in maintenance mode and sends new users to Microsoft Agent Framework.
4. Usable by one developer today. That is why Gemini CLI sits in a note rather than the main list.
5. A start path in the official README that you can run yourself.
6. Listed on AgenticMart, so you can see its licence, scan date and what it can touch before you install. Strong agents without a listing are named above, not ranked.

Stars were used to find candidates from the AgenticMart agents leaderboard, not to rank them. We did not run the agents for this post and do not rate their output quality.

Before you self-host an AI agent

Self-hosted AI agents keep your data off someone else's servers, but they run with your permissions. Four habits cover most of the risk:

• Start narrow. Give a new agent one project folder, not your home directory. Our least-privilege guide shows how.
• Use a container or VM for anything that runs shell commands. Hermes Agent, DeerFlow and DeepSeek Harness all document a sandboxed option.
• Keep approval prompts on for anything that writes, sends, deletes or pays.
• Pin versions. Install the release you read about, not whatever ships tomorrow.

The AI agent security guide goes deeper.

FAQ

What are the best open source AI agents in 2026?

By our criteria: Hermes Agent and Goose for general use, Open Interpreter for coding, DeerFlow for research, Browser Use, Nanobrowser and MagenticLite for the browser, and LangGraph, CrewAI and the OpenAI Agents SDK as frameworks. DeepSeek Harness is worth watching but is still a developer preview.

What is the best open source AI coding agent?

Among coding agents listed on AgenticMart, Open Interpreter and Goose are the strongest picks today. OpenCode, Codex CLI, OpenHands, Cline and Aider are popular open-source coding agents too, with MIT or Apache-2.0 licences, but they are not listed yet so we cannot show their scan panels.

Can I run AI agents fully self-hosted?

Yes. Every pick here runs on your own machine or server. To keep everything local you also need a local model, for example through Ollama, which Nanobrowser and others support. Otherwise the agent runs locally but sends prompts to a model provider.

Are open source AI agent frameworks free?

The frameworks here are MIT-licensed and free to use, including commercially. You still pay for model usage, and some vendors sell optional hosted services, such as CrewAI's control plane or Browser Use's cloud.

How do I find good AI agents on GitHub?

Check four things: the licence file, the date of the last commit, whether the README says the project is archived or in maintenance mode, and what the agent can reach. The AgenticMart agents directory puts these on one page per listing, and the leaderboard ranks listings by stars with a CSV you can download.

Is AutoGen still maintained?

Not for new work. Its README describes it as archived and in maintenance mode, with contributions limited to bug fixes, security patches and docs, and points new projects to Microsoft Agent Framework, with a migration guide for existing users.

Built an open-source agent?

If your agent missed this list because it is not listed, fix that. On AgenticMart you can list free now, paid selling opens soon. Each listing shows its licence, an automated scan with a date, and a "what it can touch" panel. List your agent, or see how builders present agent skills and MCP servers.