An open-source agent harness where agents are Python programs that improve themselves.
Website · Install · Documentation
English · 简体中文 · 繁體中文 · 日本語 · 한국어 · Español · Français · Deutsch · Português · Русский
- 2026-11-16 · v1.4.0 — Several keys and sign-ins per provider with automatic switching at a limit, task blocks in replies, and a run history for scheduled tasks. Releases are published from this repository from now on.
- 2026-11-15 · v1.3.0 — A Providers page with health, quota and spend; the agents working in a conversation listed in Activity; files an agent makes open on their own.
- 2026-11-14 · v1.2.0 — Branches of one conversation run side by side, Stop takes effect at once, and the Windows App arrives.
- 2026-11-10 · v1.1.0 — Rewind from any sent message; the App comes to Linux.
- 2026-11-06 · v1.0.0 — First release of the 1.x line: desktop App with split views, background computer use, meetings and an always-on scheduler.
curl -fsSL https://fuzzykuji.github.io/openprogram-site/install | shOn Windows (from v1.2.0): irm https://fuzzykuji.github.io/openprogram-site/install.ps1 | iex. The macOS desktop App is a DMG on the releases page; the Linux App is a deb or AppImage on the same page, and the Windows App arrives in v1.1.1. If the website does not answer: curl -fsSL https://raw.githubusercontent.com/Fzkuji/OpenProgram/main/docs/_static_root/install.sh | sh. More options in installation.
openprogram # setup wizard, then terminal chat
openprogram web # workspace at http://localhost:18100Next: getting started · models · all docs
from openprogram import Agent
class SupportAgent(Agent):
def triage(self, ticket: str) -> str:
"""Classify the ticket, then draft a reply."""
kind = self(ticket, choices=["bug", "question"])
if kind == "bug":
logs = search_logs(ticket) # your own helper
return self(f"Reply using:\n{logs}")
return self("Draft a short reply.")The model decides; Python makes sure the required steps run. See writing agents.
Each entry links to its design document.
| Agent classes | An agent is a Python class; branches, retries and checks are code, self(...) is a model call. |
| Execution DAG | Every user turn, model call and function call is a node you can inspect, branch and replay. |
| Context | Each call reads a selection of nodes instead of a growing transcript, with automatic compaction. |
| Self-programming | Agents write new methods into their own class; they become tools on the next turn. |
| Checked decisions | The model picks from a fixed menu and code validates the pick before anything runs. |
| Sub-agents | Sub-agents run on their own branch, optionally in a Git worktree, and return only their result. |
| Events | A typed event bus; plugins observe events or gate tool calls. |
| Always-on agents | Agents that keep running in the background and react to events. |
| Goals and recovery | Long-running goals survive restarts and resume where they stopped. |
| Fault tolerance | Retries, timeouts and error classification for every model call. |
| Desktop and web UI | The App on macOS and Linux, and from v1.2.0 on Windows; the web UI on any of them or a server. |
| Split panes | Chats, files, terminals and web pages side by side. |
| Built-in browser | A browser the agent reads and operates while you watch. |
| Integrated terminal | Terminals in the workspace, shared with the agent. |
| Computer use | Operate desktop apps, browser pages or a remote VM. |
| CLI and TUI | The same agent from the terminal. |
| Remote access | Reach your own instance from another device, owner-only. |
| Configurable Agents | Save model, instructions, tools and memory access as a reusable Agent. |
| Memory | Source-backed memory with retrieval and background writing. |
| Permissions and sandbox | Approve or deny tool calls with a note; commands run under a sandbox. |
| Programs | Tools and applications as installable Programs. |
| Providers | Anthropic, OpenAI, Gemini and more, several accounts per provider. |
| Local models | Local inference servers, with optional authentication. |
| MCP | Use MCP servers, or expose OpenProgram as one. |
| Skills and plugins | Extend what the agent knows and can call. |
| Channels | Talk to your agents from Telegram, Discord, Slack and WeChat. |
| Automatic updates | Signed releases that update in place. |
@inproceedings{qi2026llmascode,
title = {LLM-as-Code: Agentic Programming for Agent Harness},
author = {Qi, Junjia and Fu, Zichuan and Gao, Jingtong and Zhang, Wenlin and Yan, Hanyu and Wu, Xian and Zhao, Xiangyu},
booktitle = {KDD 2026 Workshop on Agentic Software Engineering (AgenticSE)},
year = {2026},
url = {https://arxiv.org/abs/2606.15874},
}AGPL-3.0 © 2026 FuzzyKuji. See the license for use, modification, distribution and network-service obligations.
Thanks to contributors who provide code, reports, reproductions, tests, documentation and reviews.