Open-Source · Agentic Workflow Harness

OpenProgram

Build your agent workflows in Python.
Any LLM · Any platform · Self-evolving.

v0.4.0AGPL-3.0Python 3.11+ macOS · Linux · WindowsOSWorld Multi-Apps 79.8%

OpenProgram · github.com/Fzkuji/OpenProgram01

The problem

An LLM is flexible. Code is deterministic.

Pick one extreme and you lose. A harness balances the two — interleaved moment to moment.

Let the model run everything

Unpredictable execution
Context explosion
No output guarantees

Python for the flow.
LLM for the judgement.

Hard-code everything

Brittle, no adaptation
You lose the intelligence
Can't handle novelty

Why a harness · Agentic Programming02

The proposal

Agentic Programming

Python drives the control flow you want fixed.
The LLM reasons only when you ask it to.

“The more constraints one imposes, the more one frees oneself.”
— Igor Stravinsky, Poetics of Music

Deterministic flowFlexible reasoning Automatic DAG contextOutput guarantees

The idea · @agentic_function03

Why OpenProgram

Five things you get out of the box

01

Deterministic flow

Python drives control flow; the LLM reasons only when asked.

02

Run it anywhere

Native macOS / Linux / Windows — terminal, browser, or chat. No WSL, no Docker.

03

Automatic context

A shared DAG threads context into every call. Multi-agent ready.

04

Any LLM, any provider

API key, or the CLI subscription you already pay for — Claude / Codex / Gemini.

05

Self-evolving

The agent builds, runs, and improves its own workflows and tools.

→

One backend

~/.openprogram — terminal and browser share sessions live.

Capabilities04

Same agent · same task · different reliability

Why agentic workflows matter

Skills optimize for flexibility · Workflows optimize for dependable execution05

Quick start

From zero to chatting in four steps

1

Install

pip install openprogram
openprogram setup

The wizard adopts any CLI you're logged into, or prompts for a key.

2

Chat

openprogram        # TUI
openprogram web    # browser

Web on :18100, API on :18109. One backend behind both.

3

Write functions

"create an @agentic_
function that …"

Ask in chat — the skill handles decorator, smoke test, validation.

4

Add harnesses

openprogram programs
install research

Clones in, pip-installs deps, auto-registers on restart.

Quick Start · pip install openprogram06

The harness suite

Three siblings, one install

Any third-party harness plugs in the same way — no symlinks, identical on Windows.

GUI Agent

Observe → plan → click → verify by vision. Drives desktop apps & OSWorld VMs across macOS / Windows / Linux.

OSWorld Multi-Apps 79.8%

Research Agent

Literature survey → idea → experiments → paper draft → cross-model review.

Topic → submission draft

Wiki Agent

Ingests notes / docs / chats into an Obsidian-compatible vault with [[wikilinks]].

Obsidian vault output

Programs · openprogram programs install gui / research / wiki07

See & steer

A conversation as a git DAG

The web UI draws a live mini-DAG of the session — branch / merge / attach on any node, multi-agent rows tagged by producer, drag-and-drop attachments.

Mini-DAG viewBranch · merge · cherry-pick Multi-agentLayered memorydeep_work loop

Web UI · openprogram web → localhost:1810008

Get started today

Python for the flow.
The LLM for the judgement.

Open-source agentic harness — install with one pip command, bring any LLM, and let your workflows evolve themselves.

github.com/Fzkuji/OpenProgram Getting Started Philosophy