OpenFlow is a visual, drag-and-drop builder for composing and running multi‑agent AI workflows. Users place role cards on a canvas, wire them into pipelines (typical flows shown: planner → architect → coder), save setups, and execute real parallel agents driven by the opencode headless engine. The project ships as a fork of opencode but keeps all OpenFlow code confined to packages/flow so the upstream engine remains unmodified. The repository includes extensive docs (AGENTS.md, FLOW.md, many localized READMEs), an MIT license, and tooling for local development and deployment.
Practical details matter: OpenFlow requires Bun 1.3+ and Git; install by cloning and running bun install. The engine listens on port 4096 and the canvas on 5174; the recommended launcher is FLOW_MANAGE_SERVER=1 bun openflow.ts which lets the canvas start, restart and re-read the engine (so applying merged agents or MCP changes is easy). Troubleshooting notes cover port conflicts and a shared database lock condition (commands shown to kill conflicting processes). A built mode, dry run, and project/port flags are supported (OPENFLOW_PROJECT, OPENCODE_SERVER_URL, OPENFLOW_BUILT, OPENFLOW_DRY_RUN). A free opencode "zen" model tier (IDs ending in -free) works without keys, and an in-app API Keys panel lets users add or import provider keys; workflows edit real files in the configured project.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.