What Is Flowfield? Shared Task Feeds for AI Agents

Flowfield is a local service that runs entirely on your machine and acts as the coordination layer between you and AI coding agents. It stores all data in ~/.flowfield/, exposes a browser UI at localhost:8765, and provides a CLI for scripting and integration. Your coordinator agent plans and captures work; worker agents implement tasks in isolated Git checkouts; and you retain direct control of priorities, questions, and every approval decision. Flowfield is not a cloud service — no account is required and no project data leaves your machine.

​

What Flowfield does

• Shared task feeds. Every task carries a single feed containing its definition, revisions, worker activity, questions, your answers, testing observations, and results. Intent and evidence stay together in one place — not scattered across chat windows, pull request comments, or separate logs.
• Parallel workers in isolated Git checkouts. Independent tasks run simultaneously, each in its own checkout. Tasks that depend on other tasks wait until their prerequisites are complete and their code is available. You control how many workers run concurrently.
• Explicit approve-before-integrate review. Workers deliver a result to a review state — not directly to your branch. You read the outcome, inspect the diff, optionally try the result in an independent copy, record testing observations, and then choose to approve and integrate or request changes. No code reaches your project without your explicit action.

​

Who it’s for

Flowfield is for developers who use AI coding agents — like Codex — to implement real work on real projects and want visibility, control, and reliable delivery. If you have ever lost track of what an agent agreed to do, wondered whether its output passed your checks, or wanted a way to review a diff before it landed in your branch, Flowfield was built for you. It is equally useful whether you run one task at a time or keep several workers busy in parallel.

​

How it works

You run flowfield serve in a terminal to start the local service. The browser UI opens at localhost:8765 and shows your board — a view of every task across its lifecycle stages: Backlog, Up Next, In Progress, In Review, and Done. You connect your Codex coordinator to Flowfield via MCP by running flowfield integration connect codex in your project directory. The next time you start a fresh Codex conversation, the coordinator loads your project’s Flowfield skill and can read the board, capture tasks, ask and answer questions, and report results — all through a structured connection, not freeform chat. You describe a piece of work to your coordinator and ask it to capture and prepare the task in Flowfield. The coordinator records the agreed definition, success criteria, and any dependencies. You set the priority by moving the task to Up Next, then click Run Queue to let the service start eligible work. Workers run the task in an isolated Git checkout, posting progress and any questions to the task feed as they go. You answer questions directly in the feed. When a worker finishes, the task moves to In Review — you read the result, inspect the diff or try the code, then approve and integrate, or request changes. Flowfield delivers approved code to your configured project checkout and marks the task Done.

​

Key concepts quick-reference