The latest articles on DEV Community by Bellows AI (@bellows). https://dev.to/bellows https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4157669%2F21e5d6cf-f4ac-49d2-87bd-0226f6c2db77.png https://dev.to/bellows en Bellows AI Mon, 05 Oct 2026 17:43:02 +0000 https://dev.to/bellows/i-use-three-coding-agents-so-i-built-one-desktop-app-to-run-them-all-under-one-set-of-rules-jll https://dev.to/bellows/i-use-three-coding-agents-so-i-built-one-desktop-app-to-run-them-all-under-one-set-of-rules-jll <p>Most days I use more than one coding agent: Codex for one job, Gemini CLI for another, Claude for a third. Each lives in its own terminal, asks permission its own way, and leaves no shared record of what it was allowed to do.</p> <p>So I built <strong>Bellows</strong>: a free desktop app for Windows, Mac and Linux that runs them all side by side, under one set of rules.</p> <h2> Every agent, one window </h2> <p>Run Codex, Gemini CLI and Claude sessions in tabs, in the same project. Gemini CLI, Codex and Qwen Code have presets, and Bellows is an <strong>Agent Client Protocol (ACP) client</strong>, so any agent that speaks ACP is one command away. If Ollama or LM Studio is running, one click starts a session on a local model.</p> <h2> One set of rules, one record </h2> <p>Permission prompts are buttons, and they look the same whichever agent asks. Save an answer as a rule and it applies to every agent. Every allow and deny goes into a <strong>hash-linked audit trail</strong>, so an edit or a gap shows.</p> <p>Codex runs through its own app-server, the one mode in which it asks before commands and edits. The few commands it treats as safe still run unasked, and Bellows records each one as having run.</p> <h2> Pair on a live session </h2> <p>Send a colleague a link to your running session. Their link carries one role: <strong>watch, suggest or type</strong>. Suggestions wait for you to accept them, and answering permission prompts is a separate switch, off by default. Everything stays on your machine. On your own network it's free.</p> <h2> Keep what worked </h2> <p>When an agent gets a file job right, press <strong>Keep as a tool</strong>. It becomes a small app with a Run button: next month, pick next month's files and press Run, with no agent and no AI charge. Kept tools run sealed off (no network, no shell), can be sent to a colleague as one file, and <strong>Check</strong> proves they still work.</p> <h2> Room to experiment </h2> <p>Start any agent in its own <strong>git worktree</strong> and review what it changed before you keep it. Resume or fork old sessions, and rewind files in Claude sessions.</p> <h2> What it doesn't do (yet) </h2> <ul> <li>No model included: each agent signs in its own way, and Claude runs on your Anthropic API key.</li> <li>The Windows download isn't code-signed yet (SmartScreen: <em>More info</em>, then <em>Run anyway</em>). Mac needs Apple silicon.</li> <li>Several agents on one job at once, with a side-by-side review, is what I'm building next.</li> </ul> <p>The desktop app is free, all of it. A paid Team plan adds one policy for every machine and a shared audit trail.</p> <p>Running agents side by side: <a href="https://bellowsai.app/guides/run-several-coding-agents-side-by-side" rel="noopener noreferrer">https://bellowsai.app/guides/run-several-coding-agents-side-by-side</a></p> <p>Which agents do you use together, and what would get you out of the terminal?</p> agents ai productivity softwaredevelopment