Blog · · 2 min read
Why BetterC0de is bring-your-own-CLI (and why we won't proxy your prompts)
Most AI coding tools want to sit between you and the model. BetterC0de doesn't. Here is the reasoning behind running Claude Code, Codex, Cursor Agent and Grok through their own CLIs.
There are two ways to build a multi-agent coding workspace. You can put yourself in the middle — take the user's key, call the model APIs, run the tools, bill for it — or you can get out of the way and drive the vendors' own tools. BetterC0de does the second thing. This post is about why.
The middleman model has three costs
Your prompts and code go through someone else. Every proxied request is another server that sees your source. Even with the best intentions, that is a bigger attack surface and a longer privacy policy.
You pay twice. A proxy needs to charge for tokens on top of the vendor's price, or lock you into its plan. You already have a Claude Max or ChatGPT subscription; you should get to use it.
You get the proxy's version of the agent. Claude Code, Codex and Cursor Agent are not just models — they are tool loops, sandboxes, permission systems and years of prompt engineering. A wrapper that re-implements them on the raw API is a worse agent, always.
What bring-your-own-CLI gives you instead
- The real agent. BetterC0de launches
claude,codex,cursor-agentorgrokexactly as the vendor ships it and speaks their structured output. Every improvement they ship, you get on the day. - Your login, your plan. No key to paste. If you are signed into a CLI in your terminal, you are signed in here.
- Nothing in between. Prompts go from your machine to the vendor. Our servers get an anonymous heartbeat and, if the app crashes, a report. Full list here.
- Easy to add agents. Supporting a new CLI is a small adapter, not a new backend. Expect the list to grow.
What we do add
The parts the vendors don't: one editor and terminal for all of them, a single inline permission layer so every agent asks the same way, worktree-aware conversations so two agents can change code in parallel, and Diff, Git and Preview views to check the work before it lands.
That is the whole product — the workspace around the agents, not a replacement for them.