Overview#

This page answers one question: what OpenProgram can do, and which page covers each capability. Capabilities come in three layers: the programming paradigm, ready-made workflows, and extension mechanisms.

The Agentic Programming paradigm#

OpenProgram is built on Agentic Programming: Python controls the flow, the LLM provides the reasoning. You decompose a task into a function call graph — nodes that need no reasoning are plain Python, nodes that need understanding / generation / judgment are defined as an ordinary Agent method and call the model via llm(...) inside the function body. Execution order, state, and retries are ordinary code you can unit-test.

• Agentic Programming guide — the learning path for writing functions and the three "choose the next step" mechanisms
• Design philosophy — what problem the paradigm solves and why it inverts control

Agentic workflows: ready-made agents#

Complete workflows written on top of the paradigm (called harnesses / agentic programs in the code), usable right after install: GUI automation, autonomous research, and a personal knowledge base. GUI, Research, and Wiki are included in every supported release. Inspect their status with openprogram programs available and list their registered functions with openprogram programs list; the functions can be triggered as tools in chat or run directly with openprogram programs run. The install command is for third-party harnesses and developer source overlays.

• Agentic workflows overview
• GUI Agent — give it one task sentence, it operates the desktop autonomously
• Research Agent — from topic selection to a submittable paper
• Wiki Agent — distills sessions into an HTML knowledge base
• Installing and writing harnesses — the install mechanism and directory contract for third-party harnesses

Extension mechanisms#

Ways to extend the agent's capabilities without writing a harness:

• Skills — the SKILL.md registry: domain knowledge and playbooks the model loads on demand

• Distill — turn a session that worked into a reusable skill or function, so the procedure survives the conversation

• Commit, push, PR — take finished work from the working tree to a reviewable pull request, with AI co-author attribution in git log

• Plugins — install plugins from pip / npm / git / local paths that contribute commands, skills, MCP servers, and more to the host

• MCP — connect any MCP server; its tools appear directly in chat

• Built-in tools — the tools that ship with the framework (shell, files, web search, images, PDF, etc.) and the keys each one needs

• Configure an Agent — models, instructions, capabilities, memory and independent trial conversations