Ten open-source AI agents you can run on your own hardware in 2026, with their GitHub repos, licenses and what the optional paid plans cost.
The best open source AI agent in 2026 depends on where you want it to work. For a personal assistant you message from WhatsApp, Telegram or Slack and that acts on your own machine, the pick is OpenClaw. If that assistant should also learn from finished jobs and run tasks on a schedule, Hermes Agent fits better. Most developers are well served by OpenCode. For desktop work beyond code, goose is the easiest start.
Six more agents cover narrower jobs, from reviewed edits in the editor to web automation and building your own agent. All ten are MIT or Apache-2.0 licensed. Each entry links the GitHub repository, then says what costs money on top.
Tools, features and prices checked on October 24, 2026.
What changed for open-source AI agents in 2026
Open source agents became finished products this year. OpenClaw, a personal agent you reach through chat apps, now has more GitHub stars than React or the Linux kernel. Hermes Agent stood at about 87,500 stars in mid-April and has about 257,000 now. Governance changed with them. OpenClaw is run by the OpenClaw Foundation, a US non-profit, while its creator Peter Steinberger works at OpenAI. Block handed goose to the Linux Foundation’s Agentic AI Foundation, announced on April 7.
Agents also started to plug into each other. ACP, a shared protocol between agents and editors, lets OpenHands run Claude Code, Codex or OpenCode from one screen. Open-weight models keep the bill low: on Nous Portal, gpt-oss-120b costs $0.03 per million input tokens. Closed tools went the other way. Google stopped serving Gemini CLI to consumer accounts on June 18 and moved them to Antigravity CLI. Security work is visible now, too: ClawScan reviews skills uploaded to OpenClaw’s ClawHub, and Tencent’s AIG joined it in October.
How we picked these tools
Each agent here ships under an OSI-approved license (MIT or Apache-2.0) and had a release in the month before our check. Eight of them run without writing code. Browser Use and Deep Agents are libraries, included for readers who automate the web or build their own agent. Source-available licenses, closed products and bare frameworks stayed off. Stars, licenses and release dates come from GitHub; features and prices from each vendor’s own site and docs.
- Where it runs: terminal, editor, desktop app, server or chat apps.
- Model freedom: any provider or a local model, instead of one vendor’s API.
- Safety controls: sandboxes, approval steps and undo, because these agents run commands on real machines.
- Memory and skills: whether context and past solutions carry over.
- Unattended work: schedules, webhooks and background runs with nobody at the keyboard.
- Cost to run: what stays free, and what models, hosting or team plans add.
1. OpenClaw
Texting your own computer is the idea behind OpenClaw. You write to it on WhatsApp, Telegram, Signal, iMessage or any other of its 29 channels, and it browses the web, fills in forms, edits files and runs shell commands, with full access or in a sandbox. It remembers preferences, projects and people. Hosted and local models both work. The October 19 release added sidebar agents and Skill Workshop 2.0.
Key AI features
- Self-written skills: the agent writes new skills for itself, and ClawHub adds community-built ones.
- Team gateway: one gateway serves a whole team with shared sessions.
Pros
- No new interface to learn, since it lives in chat apps you open every day
- Routine jobs can start from a community skill instead of a blank prompt
- No subscription and no upsell, because the foundation sells no paid tier
Cons
- Shell, file and account access widen the attack surface, so community skills need a check before install
Open source: MIT, about 393,000 stars on GitHub (openclaw/openclaw).
Pricing: Free, with no paid tier or hosted version; you pay only for the model you connect.
2. Hermes Agent
Nous Research built Hermes Agent to get better the longer it runs. When it solves a problem, it saves the method as a skill and reuses it next time. Its memory carries over between sessions and across Telegram, Discord, Slack, WhatsApp, Signal, email and the command line. Jobs can be scheduled in plain language. A morning briefing, say, then runs unattended.
Key AI features
- Subagents: isolated helpers, each with its own conversation and terminal.
- Sandbox backends: local, Docker, SSH, Singularity or Modal.
- Built-in tools: web search, browser automation, vision, images and text-to-speech.
Pros
- Recurring tasks need less prompting over time
- Routine reports arrive without anyone asking for them
- Nothing needs explaining twice when you switch channels
- Risky experiments can run on a remote machine instead of your laptop
Cons
- Desktop apps exist for macOS and Windows only; on Linux you install from the terminal
Open source: MIT, about 257,000 stars on GitHub (NousResearch/hermes-agent), v0.22.5 from October 23.
Pricing: Free with your own provider or a connected ChatGPT account. Nous Portal adds models, hosted tools and Hermes Cloud hosting: Free $0 (free models only), Plus $20/month with $22 in credits, Super $100, Ultra $200.
3. OpenCode
With about 221,000 GitHub stars, OpenCode is the most-starred coding agent on this list, and its site counts more than 16 million developers a month. It runs in the terminal, as a desktop app, in your IDE and on the web, and you can sign in with GitHub Copilot or ChatGPT Plus or Pro or start with free models that need no account.
Key AI features
- Any model: more than 75 providers through Models.dev, local models included.
- Parallel agents: several sessions work on the same project at once.
Pros
- A ChatGPT or Copilot plan you already pay for keeps working, so you don’t pay twice
- Switching models per task keeps you independent of one vendor’s price changes
- You can try it before you enter any payment details
Cons
- The FAQ says data is stored when you use the free models, so private code belongs on your own provider key
Open source: MIT, about 221,000 stars on GitHub (anomalyco/opencode), v1.19.4 from October 22.
Pricing: Free. Zen is pay-as-you-go at provider prices with no markup, from a $20 balance. Go costs $10/month and Go Plus $40/month for 30 bundled models, Kimi K3.1 Pro and GLM-5.3 among them.
4. goose
Research, writing and data analysis are as much goose’s job as code. It comes as a desktop app for macOS, Linux and Windows, a CLI and an API, and connects to more than 70 MCP extensions for databases, browsers, GitHub or Google Drive. Recipes package a working setup as a YAML file that a colleague can rerun or a CI job can call.
Key AI features
- MCP Apps: extensions can show buttons, forms and charts inside the desktop app.
- Subagents: parallel workers for jobs such as code review or file processing.
- Safety checks: prompt-injection detection, tool permissions, a sandbox mode and an adversary reviewer.
Pros
- People who avoid the terminal can start from the desktop app
- A team shares one recipe instead of everyone rebuilding the same prompts
- An existing Claude, ChatGPT or Gemini subscription can power it through ACP, without a separate API bill
Cons
- It supports 15+ model providers, far fewer than OpenCode’s 75+
Open source: Apache-2.0, about 56,000 stars on GitHub (aaif-goose/goose), v1.56.0 from October 23.
Pricing: Free; you pay your model provider directly.
5. OpenHands
When a new GitHub issue arrives, OpenHands can split it into tasks on its own, triggered by a webhook. Agent Canvas is the self-hosted control center behind this. It runs the OpenHands agent and, through ACP, Claude Code, Codex, Gemini CLI, Pi or OpenCode. Agents run locally, in Docker, on VMs, on company servers or on OpenHands Cloud. Scheduled and webhook automations connect to Slack, GitHub, Linear and Notion.
Key AI features
- Shared agent server: one server runs a team’s code reviews and dependency updates.
- Backend switching: one frontend moves between local, remote and cloud agents.
Pros
- Agents keep working on a server after you close your laptop
- Comparing coding agents needs no separate setups
- New issues turn into work without someone starting the agent
Cons
- The quick install without Docker gives the agent full access to your filesystem
- Multi-user features with RBAC and SSO start at $499/month
Open source: MIT, about 92,000 stars on GitHub (OpenHands/OpenHands), v1.33.0 from October 23.
Pricing: Agent Canvas is free. Teams costs $499/month for up to 10 users plus $50 per extra seat, capped at 5,000 sessions a month, with a 30-day trial; Enterprise on request.
6. Cline
Developers who want to see a plan before any file changes get it from Cline. Plan mode works out an approach with you. Act mode carries it out, step by step with your approval or on auto-approve, across several files with linter-aware fixes, diffs and checkpoints. It runs in VS Code, Windsurf and Cursor, and also as a CLI, desktop app and SDK.
Key AI features
- Project rules: a .clinerules file tells it your coding standards.
- MCP servers: databases, APIs and infrastructure become tools it can call.
- Headless runs: works in CI and on cron schedules.
Pros
- A bad step is reverted in one click instead of a manual git cleanup
- A wrong approach gets caught in the plan, before tokens go into code
- Local models through Ollama or LM Studio keep code on your machine
Cons
- Approving each step slows long tasks unless you turn on auto-approve
Open source: Apache-2.0, about 71,000 stars on GitHub (cline/cline).
Pricing: Free, with no subscription or seat fee; inference runs on your own keys or the Cline provider. Enterprise adds the JetBrains extension, SSO, RBAC and audit logs, priced on request.
7. Browser Use
Hermes Agent made Browser Use its default browser tool, and Browser Use’s case study on that move cites 48 to 66% fewer task tokens. The open-source core is a Python library. With it, a language model drives a real browser for logins, forms, data extraction or a booking with several steps. Browser Harness gives coding agents such as Claude Code or Codex access to your own Chrome.
Key AI features
- MCP server: other agents call the browser as a tool.
- Site skills: reusable instructions for specific websites.
Pros
- Web chores on sites without an API become repeatable jobs
- It slots into agents you already run instead of replacing them
- Larger workloads need no browser fleet of your own
Cons
- The library takes Python, and the infrastructure built to get past bot checks belongs to the paid cloud
Open source: MIT, about 119,000 stars on GitHub (browser-use/browser-use), 0.13.13 from October 15.
Pricing: The library is free. Cloud browsers cost $0.02 an hour plus proxy traffic, hosted web agents the model price plus 20%, and new accounts get $1 in credit.
8. DeerFlow
ByteDance rewrote DeerFlow from scratch for version 2.0, which reached the top of GitHub Trending on February 28, 2026. It is now a “super agent harness” for long research, coding and content jobs that coordinates subagents, keeps memory and runs code in sandboxes, with Docker as the recommended setup and BytePlus InfoQuest for web search and crawling.
Key AI features
- Subagents: parts of a long task go to separate agents.
- Extensible skills: new abilities are added as skills.
Pros
- One request can cover a research-to-report job that would otherwise take a chain of prompts
- Generated code runs in a sandbox, away from your own system
Cons
- Setup means Docker and config files, and the README’s model advice points to ByteDance’s own coding plan
Open source: MIT, about 84,000 stars on GitHub (bytedance/deer-flow), v2.1.0 from September 24.
Pricing: Free. The README recommends ByteDance’s Volcengine and BytePlus coding plan with models such as Doubao-Seed-2.0-Code, DeepSeek v3.2 and Kimi 2.5.
9. Open Interpreter
Open Interpreter has changed course. It is now a fork of OpenAI’s Codex, tuned for low-cost and open-weight models such as Kimi K3 or GLM 5.3. Its /harness command switches between emulated harnesses, including those of Claude Code, Kimi Code, Qwen Code and SWE-agent, so each model gets the prompting it works best with.
Key AI features
- Native sandboxing: commands run sandboxed on macOS, Linux and Windows.
- QA skill: any model can operate and test web and native apps.
- ACP agent: runs inside ACP-compatible editors.
Pros
- Low-cost models get the prompting that suits them, which keeps the bill per task small
- Projects on the Codex SDK switch with a one-line change
- Shared formats such as AGENTS.md make a later move to another agent easy
- Paired with an open-weight model, the whole stack is open
Cons
- A full desktop or browser interface needs the separate Interpreter Workstation project
Open source: Apache-2.0, about 69,000 stars on GitHub (openinterpreter/openinterpreter).
Pricing: Free; model usage is billed by your provider.
View Open Interpreter on AI Tower
10. LangChain Deep Agents
For readers who would rather build than adopt, LangChain offers Deep Agents, a harness on LangGraph where any piece can be extended or replaced. Out of the box it brings subagents with isolated context, a filesystem on local, sandboxed or remote backends, a sandboxed shell, persistent memory and approval of tool calls. Any model with tool calling works, local ones included, in Python or TypeScript.
Key AI features
- Context management: long threads get summarized and tool output is offloaded to disk.
- Skills and MCP: behaviors load on demand, and any MCP server can serve as a tool.
- Deep Agents Code: a ready-made coding agent for the terminal.
Pros
- You skip writing planning and memory yourself, and the defaults are tuned for long, multi-step work
- Replacing one part needs no fork, so the agent can grow with your product
Cons
- It takes code to use, and deployment through LangSmith needs the paid Plus plan
Open source: MIT, about 31,000 stars on GitHub (langchain-ai/deepagents), 0.7.32 from October 23.
Pricing: The library is free. LangSmith Developer is $0 with 5,000 traces a month, Plus $39 per seat per month with 10,000 traces and one free small deployment, Enterprise on request.
Open source AI agents compared
| Tool | Best for | Standout AI feature | Price from |
|---|---|---|---|
| OpenClaw | Assistant in chat apps | 29 channels, self-written skills | Free |
| Hermes Agent | Agent that learns | Reusable skills, schedules | Free; Portal $20/month |
| OpenCode | Coding with any model | 75+ providers | Free; Go $10/month |
| goose | Desktop work beyond code | 70+ MCP extensions | Free |
| OpenHands | Agents on team servers | Webhook automations | Free; Teams $499/month |
| Cline | Reviewed edits | Plan/Act, checkpoints | Free |
| Browser Use | Web tasks | LLM-driven browser | Free; cloud $0.02/hour |
| DeerFlow | Long research jobs | Subagents in sandboxes | Free |
| Open Interpreter | Cheap open models | Per-model harness | Free |
| LangChain Deep Agents | Building your own | Extensible harness | Free; LangSmith $39/seat |
What didn’t make the list
AutoGPT’s platform code is under the PolyForm Shield license and Kortix, formerly Suna, under Elastic 2.0, so both are source-available. n8n uses its own Sustainable Use License and is a workflow tool first. Manus is closed, and Claude Code’s license reserves all rights to Anthropic. Codex CLI and Pi overlap with OpenCode and with Open Interpreter, which is built on Codex. CrewAI, LangGraph and AutoGen are frameworks; Deep Agents covers that slot.
FAQ
What is the difference between an open-source AI agent and an agent framework?
An agent works on tasks as soon as you install it, while a framework is a library you write code against. OpenClaw and OpenCode are agents; CrewAI and LangGraph are frameworks. Deep Agents sits in between, since the library ships a working default agent and a coding agent for the terminal.
Are open-source AI agents really free?
The software is free, but the model behind it usually costs money. You pay a provider per token, reuse a subscription or run a local model on your own hardware. Paid tiers from the projects sell convenience: hosted models, bundled model plans, cloud browsers or team management.
Can open-source AI agents run on local models?
Yes, most agents on this list accept local models. OpenCode and OpenClaw support them directly, goose connects to Ollama, and Cline names Ollama and LM Studio. Deep Agents works with any model that supports tool calling, and Open Interpreter adapts its prompting to the open model you pick.
Is it safe to give an AI agent access to my computer?
It is reasonably safe once you set limits. Run unattended agents in a sandbox, such as Hermes Agent’s Docker backend or goose’s sandbox mode. Keep step approvals on for anything that touches important files, as Cline allows, and check community skills before you install them.
Which open-source AI agent works with WhatsApp or Telegram?
OpenClaw and Hermes Agent both do. OpenClaw covers 29 channels, WhatsApp, Telegram and iMessage among them, in direct messages and groups. Hermes Agent answers on Telegram, WhatsApp, Signal, Slack, Discord and email with one shared memory. Cline connects to Telegram too, though it is built around coding.
Verdict
Pick by where the agent should live. In your chat apps, OpenClaw reaches you on more channels, while Hermes Agent is the one that learns and works to a schedule. In the editor or terminal, OpenCode covers most needs and Cline adds tighter control. Teams with their own servers should look at OpenHands. goose handles desktop work, Browser Use websites and DeerFlow long research runs. Deep Agents is where builders start.
