Roxels/ docs
mcp

Claude Code skill: /run-roxels

/run-roxels is a Claude Code skill that turns a Roxels voice call into instructions for your coding session. You type /run-roxels, a conversation opens in your browser, you talk (and optionally screen-share) to dictate what you want, and when you hang up, Claude reads the transcript, plays back the instructions it heard, and acts on them.

Under the hood it runs the conversation against a default template curated by Roxels, then polls the results — you don't configure any template.

Install

Paste this one line into Claude Code (or send it to a teammate):

follow https://app.roxels.ai/skills/run-roxels/install.md

Claude fetches the guided installer, downloads the skill files, asks you three quick questions (your Roxels API key, the language the agent speaks, and whether prompts come back for your approval), writes the config, and offers the optional instant-launch hook. Restart Claude Code once and type /run-roxels.

The script is plain Python (standard library only) — no virtualenv, no dependencies. Prefer to do it by hand? The manual steps are at /skills/run-roxels/install.txt.

What you need

  • A Roxels account — on first run the skill opens a browser to sign in (or sign up for free) and click Connect Claude Code; it then provisions an API key named "Claude Code / run-roxels" and saves it for you. No copy-paste, no digging through Settings. (You can still drop a key into the .env by hand if you prefer — see API keys → Settings.)
  • Optionally, a language — set ROXELS_LANGUAGE=en (or pt, es) in the .env and the agent speaks that language on your calls. Omitted, the template decides.
  • That's it. Conversations and results stay in your organization.

How results come back

While you talk, the agent composes your prompt live — you can watch it build in the room, grounded in what you're screen-sharing. When you hang up, the skill feeds that composed prompt into your Claude Code session, where by default it comes back for your review: an editable box with the prompt prefilled — click "Send as my message" and it enters the chat as your message, which Claude then executes exactly as if you had typed it. Prefer zero friction? Set ROXELS_REQUIRE_APPROVAL=false in the skill's .env and Claude executes the prompt directly. For templates that don't stream a prompt, the skill falls back to the full transcript, which Claude distills first. The full structured payload — prompt blocks, transcript, extracted data, recording link — is saved as JSON next to where you ran it.

Last reviewed 2026-07-08.