Cukii Chat puts the command-line agents you use — Claude Code, Codex, Qwen, Grok, Cursor, Kimi — behind a single chat panel inside VS Code. No per-vendor terminal juggling and no second UI to learn: open the panel, install the CLI you already pay for, sign in, and talk to it.

What you get

• One composer, many vendors. Switch between Claude, OpenAI Codex, Alibaba Qwen, xAI Grok, Cursor and Moonshot Kimi from the model picker, with a Best / All toggle and a per-session effort control.
• Your own accounts. Manage accounts installs a vendor CLI, then you sign in with the account you already pay for. Qwen asks you to paste an API key. Each vendor shows one connected / not-connected dot.
• Talk while it works. A message you send mid-run reaches the agent at its next step, with read receipts on your bubble. The agent can ask you a question in a shared dialog, laugh at your joke and react to your message — without dropping the task.
• Sessions name themselves. The agent titles a new chat after your task; a name you set yourself always wins. Account & usage shows your plan's session and weekly limits wherever the vendor reports them.
• Sessions that survive. Interrupted runs, remote SSH windows and restored history come back where you left them.
• Long answers stay readable. Sticky user messages fold to a single line while you scroll, and unfold on demand.
• Bug reports on GitHub. Report a stuck or wrong session with screenshots and masked diagnostics. Use your GitHub account; no tracker invitation is needed.
• Wait for quota resets. Milky sleep is on by default in each chat. It shows the vendor's reset date, or the next check when no date is reported, and resumes the same task. Stop, sending another message, changing models or turning it off cancels the wait. Cukii checks again when VS Code reopens.
• Remote-SSH ready. The extension runs on the remote host, so the agents work on the machine that has your code.

Getting started

1. Install the extension, trust the folder when VS Code asks, and open Cukii Chat from the activity bar. The command is Cukii Chat: Open Cukii Chat in a New Tab.
2. Open Commands and model → Manage accounts…. For a vendor you already pay for, press Install, then Log in.
3. Pick a model and send your first message with Enter.

Remote-SSH. Run Install and Log in in the remote window. The CLI is installed in that remote user's home and runs there. A copy that exists only on your laptop is not used. The sign-in page still opens on your local computer.

If the page does not open, allow VS Code's prompt to open a website, or copy the https link from the terminal Cukii · <vendor> login. Codex prints that link in the Output channel Cukii Chat · codex login and puts the one-time code on the clipboard. For Qwen, paste an sk- key from the Alibaba console. The account row then shows connected.

Requirements

• VS Code 1.70 or newer, and a trusted workspace
• One vendor account you already pay for. Install downloads that CLI
• macOS and Linux: curl and tar. Install downloads Node.js. You do not install Node.js yourself
• Windows: Install uses npm when it is already on PATH. Otherwise it installs the current Node.js LTS with winget and may ask for administrator approval. Kimi on Windows and Cursor use those vendors' own installers

Privacy

Cukii Chat sends no analytics or telemetry. It talks to the network only for these:

• Your vendor CLIs. Your prompts go from the official vendor CLIs (Claude Code, Codex, Qwen Code, Grok, Cursor, Kimi) to their vendors, under your own accounts and their terms. Cukii Chat runs the unmodified CLIs and does not proxy your traffic.
• Usage limits. Claude and Codex limits come from the CLIs themselves; your stored Claude or Codex login is not read unless you turn on cukii.usage.directVendorApi. Kimi and Grok limits come from those vendors' own usage endpoints, signed in with the login their CLI keeps.
• Bug reports. Signing in does not publish a report. Review the text and files, then press Send. Send publishes them publicly: your description, severity, the chat snapshot unless you remove it, up to three images, and masked diagnostics. Those files are public. Screenshots are not masked. The GitHub token is used only to publish the issue and is not stored. Older private reports stay in a separate local archive.
• Updates. The vendor CLIs that Cukii Chat installed for you are updated from npm, and the model catalogue is refreshed from this repository.
• Memory. Only after you connect a Cukii Box memory endpoint with your own token.

Feedback

Open Commands and model → Report an issue…. On the first report, choose Sign in to GitHub. VS Code signs you in, including over Remote-SSH, with the scopes public_repo and read:user. No personal access token or team invitation is required. Signing in does not publish. Review the description, the chat snapshot, any images and Automatic diagnostics. Pressing Send publishes them publicly in the public tracker, with those files linked from the issue. A queued report stays on this machine and retries until GitHub confirms delivery.

License

Apache-2.0 — see LICENSE and NOTICE.