A wave of free, open-source AI agent tools shipped in the first half of 2026. Not demos — real software from Block, Alibaba, Nous Research, and independent builders, with Apache and MIT licenses, production user counts in the thousands, and GitHub stars climbing past five digits. This roundup covers thirteen of them, grouped by the job they do: chat, orchestration, parallel coding, browser sharing, code review, and mobile. Every entry is free, every entry is open-source, and every entry links to a real repository you can clone today.
If you are building a free AI agent operating system this year, these are the components worth bolting in.
Buzz (Block) — Slack for AI agents
Buzz is Block's open-source workspace where humans and AI agents share the same channels. Jack Dorsey announced it on July 21, 2026, framed as a tool "built to reduce our dependency on Slack and GitHub." It runs on Nostr — every participant, human or agent, holds a cryptographic identity, and every message, code patch, and workflow step lands in one signed, auditable event log.
Inside Buzz, agents can search prior discussions, open repositories, submit patches, review code, run YAML-based workflows, and create channels on their own. The Git forge uses Git Smart HTTP, so a feature branch becomes its own channel with patches, CI results, and merge decisions preserved alongside the conversation.
It is Apache 2.0, self-hostable, and works with Goose, Codex, and Claude Code harnesses. Caveat: by Block's own admission it has "rough edges and giant chasms." No iOS app yet. Treat it as a side-by-side experiment, not a Slack replacement.
Paperclip — agent orchestration for a whole company
Paperclip is an MIT-licensed Node.js server with a React UI that orchestrates a team of AI agents to run a business. The framing is blunt: "If OpenClaw is an employee, Paperclip is the company."
You define a goal (say, "build the #1 AI note-taking app to $1M MRR"), assign agents as a CEO, CTO, engineers, designers, marketers — any bot, any provider — and hit go. Under the hood it handles org charts, budgets, governance, goal alignment, execution policies with approval gates, cost tracking per agent and project, and cron/webhook-triggered routines.
It works with Claude Code, Codex, Cursor, Bash agents, and HTTP bots. Atomic task checkout prevents double-work. Persistent agent state lets agents resume task context across heartbeats instead of restarting. Budget hard-stops pause agents and cancel queued work when spend exceeds limits. If you have ever burned hundreds of dollars on runaway agent loops, that last feature alone is worth the install.
Orca (stablyai) — parallel coding agents in one IDE
Orca is an MIT-licensed "agent development environment" for working with a fleet of parallel coding agents. One prompt fans out across five agents, each in its own isolated git worktree — you compare the results and merge the winner.
Beyond parallel worktrees, Orca ships Ghostty-class terminals with WebGL rendering, a Design Mode where you click any UI element in a real Chromium window to send its HTML, CSS, and a cropped screenshot straight into your agent's prompt, native GitHub and Linear integration, SSH worktrees, and usage tracking that shows Claude and Codex rate-limit resets so you can hot-swap accounts mid-session.
It works with any CLI agent — Claude Code, Codex, Grok, Cursor, GitHub Copilot, OpenCode, MiMo Code, Amp, and others. 34.8K stars. Available on desktop, mobile, and VPS.
Hermes Agent (Nous Research) — the self-improving agent
Hermes Agent is the open-source MIT agent framework built by Nous Research — and yes, this site runs on it. What makes it different from every other agent on this list is the learning loop: it creates skills from experience, improves them during use, nudges itself to persist knowledge, searches its own past conversations, and builds a deepening model of who you are across sessions.
It ships with persistent FTS5-backed memory, 40+ built-in tools, a skills system that stores procedural memory as portable shareable files, and gateways for terminal, Telegram, Discord, Slack, and WhatsApp. Six terminal backends (local, Docker, SSH, Daytona, Singularity, Modal) let you run it from a $5 VPS up to enterprise serverless. 32K+ stars.
For builders who want a persistent memory system without rolling their own, Hermes ships one in the box. For tips on configuration that actually save time, see our Hermes Agent pro tips guide.
ego (lite) — a browser you share with your agents
ego (lite) is an MIT-licensed browser built by CitroLabs where you and your AI agents work in parallel. The core insight: existing browser automation frameworks (browser-use, Vercel's agent-browser) need a separate browser to drive, logins never carry cleanly, and you and the agent end up fighting for the same tabs. ego lite is one browser designed for the two of you to share.
Agents run multiple browser tasks in their own Spaces while your tabs stay yours. On first launch ego asks whether to migrate your Chrome data — say yes and your agent inherits your existing logins, cookies, extensions, and bookmarks. No API keys, no re-authentication, no Cloudflare challenges from a sandboxed fingerprint.
It ships an ego-browser skill for Claude Code, Cursor, and Codex. Benchmarks against Vercel's agent-browser show ego lite finishing tasks up to 2.5× faster with fewer tokens. macOS only today; Windows and Linux are on the roadmap.
Open Code Review (Alibaba) — deterministic AI PR review
Open Code Review (CLI: ocr) is Alibaba's Apache-2.0 AI code review tool, battle-tested internally for two years by over 20,000 developers and deployed to catch over one million code defects before it was open-sourced in June 2026.
The interesting design choice is the hybrid architecture. Most AI code review pipelines dump a diff into an LLM and hope. Alibaba built a deterministic layer first — it decides which files must be reviewed, groups related files into bundles, matches rules per file, runs extra positioning and reflection passes on comments — and only then lets the model reason inside that boundary. The result: better file coverage on large PRs, line-accurate comments instead of drift, and stable quality instead of prompt roulette.
It ships a built-in fine-tuned ruleset (NPE, thread-safety, XSS, SQL injection), OpenAI and Anthropic-compatible LLM endpoints, a VS Code extension, and skill/plugin paths for Claude Code and Codex. 17K stars. If your current AI review bot feels brittle, this is the cleaner model.
T3 Code — a GUI for managing your coding agents
T3 Code is a free MIT-licensed desktop app built by Theo and the T3 team. It is not a model — it is a front-end that connects to OpenAI's Codex or Anthropic's Claude Code using your own API keys, giving you a graphical interface for interactions that would otherwise happen in a terminal.
Git worktrees give each agent isolated workspace. You commit, push, and open a pull request in one action. A diff viewer and integrated terminal keep you inside the app. Custom quick actions let you encode repetitive workflows. Performance is the selling point — the T3 team built it to be fast on large repos.
Install via winget install T3Tools.T3Code on Windows, brew install --cask t3-code on macOS, or yay -S t3code-bin on Arch. The team is upfront: "We are very very early in this project. Expect bugs."
jcode — terminal coding agent with SSH reach
jcode is an open-source terminal coding agent that works where your code lives — including remote servers over SSH. It reads your codebase, writes surgical edits, runs commands, and shows every step. The pitch is no black boxes: local, browser, or native desktop app, and remote boxes over SSH or Docker.
It is built to keep prompt overhead small, which matters when you are paying per token. If you want to cut Claude Code tokens by up to 80%, pairing a lean agent like jcode with disciplined prompting is one of the cheaper paths.
The rest of the stack: book-to-skill, ADHD, World Monitor, and more
The video rounds out with a handful of smaller tools worth knowing about even if they are not yet headline projects:
- Book to Skill — utilities that turn a documentation page or long-form reference into a portable agent skill file, so proven workflows become reusable procedural memory instead of one-shot prompts.
- ADHD — a verbosity reducer that strips maximalist padding from agent outputs so you get the decision, not the essay. Useful when an agent is technically correct but unusably long.
- Pascal Editor — a 3D scene-building editor for agents that need to reason about spatial or physical layouts.
- World Monitor — a dashboard for watching multiple agent sessions and environments from one window so nothing drifts unattended.
- MattPCO skills — a community skill pack for agentic PC orchestration plumbing.
These are early, niche, or single-author projects — try them when you hit the specific problem they solve, not as daily drivers.
Which one should you actually install?
Pick by the job, not by the star count:
- You want agents and humans in the same chat: Buzz.
- You want to run a multi-agent company with budgets and approvals: Paperclip.
- You want to fan one coding task across five agents and merge the winner: Orca.
- You want a personal agent that remembers you across sessions and improves over time: Hermes Agent.
- You want your agent to use your logged-in browser without stealing your tabs: ego (lite).
- You want deterministic AI code review with line-accurate comments: Open Code Review.
- You want a GUI instead of a terminal for your coding agents: T3 Code.
- You want a lean terminal agent with SSH reach: jcode.
For a deeper dive into the OS-level stack these tools slot into, read how to build a free AI agent operating system. For orchestration patterns when one model is not enough, see how multi-agent orchestration beats single-model AI. For keeping agent serving costs sane at scale, see how MiniMax M3 and sparse attention reshape LLM infra.
FAQ
Are these AI agent tools actually free?
Yes. Every tool listed is either MIT or Apache-2.0 licensed, meaning you can use them commercially, modify them, and self-host without paying a license fee. What you do pay for is the underlying LLM API calls (OpenAI, Anthropic, etc.) — except for tools like Hermes Agent that can run on free inference tiers or local models.
Can I run these AI agent tools on my own server?
All thirteen are self-hostable. Buzz, Paperclip, Orca, Open Code Review, Hermes Agent, and T3 Code all ship Docker images or source builds. ego (lite) is macOS-only today. Hermes Agent runs on a $5 VPS through six terminal backends including Docker, SSH, and Modal.
What is the difference between an AI agent framework and an AI agent orchestration platform?
An agent framework (Hermes Agent, jcode) runs a single agent. An orchestration platform (Paperclip, Orca) coordinates multiple agents — assigning tasks, tracking budgets, governing approvals, and merging outputs. The article groups them by that distinction.
Which free AI agent tool is best for coding?
For pure coding, Orca (parallel worktrees, five agents at once) or jcode (lean, SSH-friendly) are the strongest 2026 picks. T3 Code adds a GUI if you prefer clicking over typing. Open Code Review handles the review side, not the writing side.
Do AI agent tools work with Claude Code and OpenAI Codex?
Most do. Orca, T3 Code, Paperclip, Buzz, and ego (lite) all explicitly support Claude Code and Codex. Hermes Agent supports them via its coding-agent skill packs. Open Code Review ships skill and plugin paths for both Claude Code and Codex.
Discussion
0 comments