GUI Agent
- Give it a natural-language task. It screenshots, clicks, types, and checks the result until the work is done.
- Same loop on the local desktop or a remote VM.
- Ships in every supported release.
Self-programming AI agent framework
Build agents that
Build agents that
program their own workflows.
OpenProgram lets agents create, run, and refine agentic functions while its runtime manages models, tools, memory, context, and multi-agent collaboration.
Start where you need
Use an agent or build your own.
Start from a shipped agent, or use the same OpenProgram runtime to create one in ordinary Python.
Research Agent
Literature → experiment → paper in one pipeline. Citations are checked against four indexes. Numbers in the paper have to match a run_record.
Wiki Agent
Pour sessions and notes into statically hostable HTML pages. Named slots accumulate; full-text search is built in.
Your agent, ordinary Python
Define an Agent class with ordinary methods. Context follows its calls automatically.
from openprogram import Agent
class Triage(Agent):
instructions = "Classify tickets as bug or feature."
def run(self, ticket: str):
return self(ticket, tools=[])Quick Start
Install, then run.
Start the Web UI at http://localhost:18100. The first run guides you through connecting a model provider.
Platform and developer installation →Included framework
Everything you need to run agents.
Use the complete runtime or adopt only the components required by your workflow. Configuration and execution remain available from Python, the Web UI, and the terminal.
providersModel providers
Connect Anthropic, OpenAI, Gemini, and compatible backends with account and model configuration.
tools + mcpTools and MCP
Give agents local tools, browser actions, external services, and MCP server capabilities.
skills + memorySkills and memory
Load reusable instructions and retain task knowledge across longer-running work.
multi-agentAgent collaboration
Spawn agents, exchange messages, and isolate file-changing work in separate branches.
sessionsSessions and branches
Keep conversations, calls, context, forks, and execution history inspectable.
web · tui · cliThree interfaces
Run the same agents from the browser, terminal UI, or headless command line.
channelsChannel integrations
Connect agent workers to supported messaging channels without replacing the runtime.
programsInstallable harnesses
Install, manage, run, and publish Python harnesses through one package contract.
One runtime, several surfaces
Use the interface that fits the task.
The Web UI and built-in Browser share sessions, providers, tools, skills, memory, and the same execution DAG.
Web UIBrowserCLIProvidersMCPSkillsMemoryChannels
Framework internals
Inspectable context, control, and events.
OpenProgram keeps agent execution inside Python while recording model calls, context dependencies, validation, and runtime events.
DAG Context
Every user turn, model call, and function call is a node. Dependency edges select the context used by branches, sub-agents, and cross-session messages.
Read the context documentation →Agentic Workflow
Python fixes required execution order, validates model decisions, retries failed checks, and combines agentic functions with ordinary tools.
Explore runtime features →Event Infrastructure
A process-wide event bus carries typed activity from the agent loop, authentication, context, channels, and memory.
Inspect the event design →Research
Related papers.
Research related to OpenProgram and its agent runtime.
Install
Build your next agent with OpenProgram.
Install the framework, connect a model provider, and start from an existing harness or your own Python workflow.