Open Source · Ready to run · Updated daily via GitHub Actions
A curated index of ready-to-run open-source agents: coding agents, browser and computer-use agents, research agents and personal assistants, with live GitHub stats. Looking for the libraries to build your own? See the frameworks index.
An AI agent is a complete program that pursues a goal by repeatedly asking an LLM what to do next and acting on the answer — reading files, running commands, browsing the web, calling APIs — until the task is done or it needs a human. The projects in this index are agents you install and run. A framework is what you would use to build one.
Most agents here follow the same loop but differ in where they live (a terminal, an IDE, a browser, a sandbox, a messaging app), what they are allowed to touch, and how they are kept in bounds. That last part — permissions, approvals, budgets, verification — is what a harness does, and it is the main thing worth comparing when you choose one.
Every agent above is tagged by what it does and where it runs. Use the Topic filter to isolate any of these — or read on for what each one means.
Agents whose main tool is a code editor and a shell: they read a repository, make changes, run tests and iterate.
Run from a command line in your project directory. Usually the most direct access to your files and shell.
Live inside an editor as an extension or a full editor. Changes show up as diffs you can accept or reject in place.
Manage many coding agents at once — parallel sessions, isolated worktrees, supervision and merge.
General-purpose assistants that live across chat apps and devices and take on everyday tasks, not just code.
Operate a web browser or a full desktop by looking at the screen and clicking, typing and scrolling like a person.
Break a question into searches, read many sources and produce a cited report rather than a single answer.
Agents built for authorized security testing. Run them only against systems you own or have written permission to test.
Domain-specific agent systems for market analysis and trading research. Treat outputs as research, not advice.
Describe themselves as a harness: they wrap the model with tools, memory, permissions and control rather than just a prompt loop.
These three get mixed up constantly, but they sit at different layers.
A finished program you run to get work done — for example a coding agent in your terminal. You configure it; you do not program it.
A library for building agents of your own: tool calling, state and an execution loop. See the frameworks index.
The control layer around an agent — what it may do, what it believes, what it can spend, and how its work is checked. Some agents ship their own; a harness can also wrap one you already use.
Six questions worth answering before you hand an agent real work, roughly in the order they eliminate options.
Terminal, IDE, browser, messaging app or a cloud sandbox? Pick the surface you already work in — the best agent you never open is worth less than a good one you do.
Some agents are tied to one vendor’s models; others take any provider or a local model. This decides cost, privacy and how easily you can switch later.
Does it ask before running commands or writing files? Can you allow-list actions? An agent that acts without asking is fast right up until it is wrong.
Does it run in a container or an isolated worktree, or directly on your machine with your credentials? Isolation is the difference between a mistake and an incident.
MIT and Apache-2.0 are unrestricted for commercial use; copyleft and source-available licenses carry obligations. Check what the tool sends home, especially in a company setting.
Agent projects move fast and some stall. The Updated column shows the last push — a popular repository that stopped moving months ago is a risk.
A program that uses an LLM to decide its own next step and then carries it out with tools — reading and writing files, running commands, browsing, calling APIs — looping until the task is done or it needs your input.
The terms overlap. “Assistant” usually means something you talk to that helps with tasks in conversation; “agent” stresses that it takes actions on its own across several steps. Many products here are both.
It depends on the agent’s permission model and where it runs. Prefer ones that ask before running commands, run inside a container or worktree, and keep credentials out of reach. Read what an agent is allowed to do before pointing it at a repository that matters.
Usually you need access to a model, either an API key, a subscription the agent can use, or a local model. Model-agnostic agents let you choose; vendor agents are tied to that vendor.
A tool for running several coding agents at once — for example in parallel branches or worktrees — and supervising them from one place. It sits above the agents rather than replacing them.
By hand: open source, ready to run rather than a library, actively maintained, and verified against the live GitHub API when the list is reviewed. Archived and long-stale projects are removed. To suggest an addition, open an issue or pull request on the buildaharness-pages repository.
Stars, forks and topics are refreshed daily via GitHub Actions — the freshness indicator in the nav bar shows the age of the current data pull. The list itself is reviewed by hand periodically.