OpenFlow overview
OpenFlow is a desktop workspace for agents. Use a direct chat for one task in a conversation; build a workflow when several agents should pass results to one another.
For example, connect a planner to a checklist agent. The planner turns your request into a plan; the checklist agent waits for it, then turns it into actionable items. You can inspect both conversations and reuse the workflow with a different request.
Start using OpenFlow
- Install OpenFlow. The packaged macOS app needs no development toolchain.
- Connect a provider: hosted credentials, ChatGPT sign-in, AWS credentials, or a local model server.
- Start a direct chat, or build your first workflow from two prompts you paste into the editor.
How OpenFlow works explains dependencies, handoffs, and your role during a run. Using the app locates the controls. For new users keeps old links working.
Work with the app
| Task | Guide |
|---|---|
| Chat with one agent | Direct chats |
| Create, configure, and validate a graph | Workflow authoring |
| Answer, approve, retry, replay, or resume | Runs and recovery |
| Work in a repository and manage saved files | Projects and persistence |
| Configure tool approval, skills, subagents, search, or MCP | Tools and integrations |
| Run a saved workflow at a time or interval | Scheduling workflows |
| Recover from a setup or run failure | Troubleshooting |
Look up behavior
| Question | Reference |
|---|---|
| How do workflows and runs relate? | Concepts |
| What does a workflow, node, edge, or run mean? | Workflows and runs |
| What does a code term mean? | Glossary |
| What are the settings, defaults, and limits? | Configuration reference |
| Which endpoints, credentials, and defaults does each provider use? | Provider profiles |
| Which tools can an agent call? | Tools reference |
| Where does OpenFlow store files and credentials? | Data locations |
| What leaves the app, and what requires trust? | Security and data |
Develop OpenFlow
Build from source sets up a development checkout. Contributing describes the change and review workflow; Architecture maps the implementation.
Contribute
| Page | Covers |
|---|---|
| Development commands | Build, check, and test commands |
| Testing workflows | Which checks a change needs |
| Development lanes | Classifying a change |
| Coding patterns | Ownership, errors, persistence, test placement |
| Build and test troubleshooting | Source-checkout failures |
| Releasing OpenFlow (macOS) | Tagging and publishing a release |
| Documentation website | The site and Markdown exports |
| README workflow demo GIF | Recording the README demo |
Architecture pages
| Page | Covers |
|---|---|
| Architecture contract | Layer boundaries and dependencies |
| Run lifecycle | UI → desktop → orchestration → engine → provider |
| Agent execution and context | Node scheduling and context |
| Orchestration crate layout | Module map |
| Provider adapters | Model transport and routing |
| Run persistence | Checkpoints, resume, replay |
| Threading and concurrency | Async runtimes |
| Callable agents in orchestration | Saved-agent snapshots |
| Overseer output repair | Repairing malformed node output |
| Plan → Execute internals | The planning gate |
| Architecture diagrams | Layer diagrams |
The repository map locates source packages. Historical plans and roadmap entries do not establish current product behavior.