What you’ll do
Point an existing coding-agent install at Zro’s private inference endpoint instead of its default provider, so agent traffic runs against open-weight models (MiniMax M3, GLM-5.2) on Zro’s EU infrastructure with zero request retention. No changes to your agent’s own config files are needed — Zro’s CLI handles session setup for you.
This matters most if you’ve been holding off on AI coding agents over where prompts and code end up: Zro’s pitch is that you get the convenience of a hosted API without your code training someone else’s model or sitting on a server outside the EU.
Step 1: Install the Zro CLI
npm install -g @moonmath-ai/zroThis installs the zro command globally, which manages authentication and launches supported coding harnesses with temporary session config pointed at Zro’s endpoint. If the global install fails with an EACCES permission error, that’s an npm global-prefix issue unrelated to Zro itself — either re-run with a Node version manager active (nvm, fnm) or fix npm’s global directory permissions rather than reaching for sudo.
Step 2: Log in
zro loginThis opens a browser flow to authenticate against your Zro account (create one at zro.moonmath.ai/sign-up first if you don’t have one) and stores a local session token. You only need to do this once — zro launch reuses the saved session afterward.
If you’re on a remote machine or a container without a browser, the CLI will print a URL and code instead of opening a window automatically — copy that URL into a browser on any device to complete the login.
Step 3: Launch your coding agent through Zro
zro launch claudeThis starts Claude Code with its model endpoint temporarily redirected to Zro instead of Anthropic’s default API, using Zro’s Anthropic-compatible /v1/messages endpoint under the hood. To pin a specific open model rather than Zro’s default routing, pass --model:
zro launch claude --model minimax-m3The same pattern works for Codex CLI and Zro’s other supported harnesses (OpenCode, Hermes, OpenClaw, Pi):
zro launch codex --model glm-5.2Step 4: Confirm it’s working
Run a normal prompt inside the launched session. Because zro launch only changes the session’s endpoint config for that run, closing the terminal or launching the tool again without the zro launch wrapper reverts to your agent’s default provider — nothing is changed permanently on disk.
To double-check traffic is actually routing through Zro rather than silently falling back to your agent’s default provider, check your usage on the Zro dashboard after sending a prompt — token counts should tick up there within a few seconds of the request completing.
Troubleshooting
zro launch claudestarts Claude Code, but it still seems to be hitting Anthropic’s default API — confirm the session actually authenticated by re-runningzro login; a silently expired token is the most common cause ofzro launchfailing open to the default provider instead of erroring.--model minimax-m3or--model glm-5.2errors with an unrecognized model name — model identifiers can change as Zro adds or renames supported models; check the pricing page for the current list rather than assuming these two names are permanent.- Session works, then stops working later in the day — session tokens expire; a plain
zro loginre-authenticates without needing to reinstall anything. - Usage on the dashboard doesn’t match what you expected — the Pro and Max plans meter by token volume, not by request count, so a handful of long-context sessions can use more of your monthly allowance than many short ones.
Where to go from here
If your tool isn’t one of Zro’s launch-supported harnesses, check the integrations page for manual setup instructions — Cursor and Cline, for example, are configured directly with an OpenAI-compatible base URL and API key rather than through zro launch. The API key for that manual path is generated from your account dashboard, the same place that shows the usage numbers referenced above. Zro’s pricing page covers plan limits if you want to check usage before pointing a high-volume workflow at it.