Voice coding with Claude Code: what Qwen Audio Agent adds, and what it costs
Updated
Illustrative photo
A voice layer on top of Claude Code
The walkthrough connects Qwen Audio Agent to Claude Code so a voice conversation can continue while the coding work runs in the background. According to the source, Qwen Audio Agent is an open-source app with browser, terminal and desktop clients. The voice front end handles the conversation, Claude Code does the coding, and a local gateway links them through an ACP adapter. The source also states that this does not make Claude write code any faster.
The source describes the setup as npm install -g qwen-audio-agent@2.0.1, a DashScope key in config.env, qwenaudio install claude, and qwenaudio setup --backend claude. The DashScope key covers the realtime voice model, and the walkthrough then starts the gateway and the browser interface. One trap the source flags: opening a terminal in a folder does not set the agent's workspace, so the workspace path has to be set in the config and checked.
Talking while the build runs
In the demo, a small issue tracker called VoiceDesk is built by voice. The source says you can ask questions whileClaude keeps building, interrupt the spoken answer, change a requirement mid-build and report a whitespace-title bug. Two limits are stated: interrupting the spoken answer stops the voice, not a task the backend has already accepted, and cancelling a task is different from muting the mic because cancelling does not undo edits already made. The source also warns that a spoken confirmation is not proof, so the app should be opened, the diff checked and the tests run.
Where it fits, and the bill
The source says the setup is great for talking through features and describing bugs while the agent works, while the keyboard should be kept for exact paths, code and errors. On cost, the software is open source, but the DashScope voice model and Claude usage are both billed separately, so the voice layer adds a second meter on top of what is already paid for Claude.
What to check before you rely on it
Before relying on a voice-driven build, the source's checks point to setting and verifying the workspace path, running a first voice test to confirm the Claude connection, and treating spoken confirmations as claims to check against the diff and the tests. The walkthrough also covers troubleshooting tips and a final verdict. The version pinned in the setup is 2.0.1, so that is the starting point the source documents.
Affiliate links: we may earn a commission from qualifying purchases, and it never changes our data or rankings. They open Amazon's search results: prices and availability are shown there.