ZeroThink CLI

The CLI now answers as ZeroThink CLI aka Agent Zero, not as a generic AI assistant. If someone asks who it is or what it can do, it answers from the local terminal runtime and current route instead of letting the model improvise.

When people want to know whether they are on Groq, NVIDIA, linked-account mode, or direct-key mode, use the runtime commands instead of asking the model to guess.

Installs the Python client, creates launchers, adds the user bin path if needed, and then prints the beginner quickstart commands.

Downloads the CLI into ~/.zerothink, creates a launcher in ~/.local/bin, and prints the same beginner next steps.

If you know nothing yet, start the guided setup before trying random commands or switching providers.

If ZeroThink CLI is already installed, this is the command to tell people to run. It downloads the latest published CLI, verifies the SHA-256 hash, then replaces the local script safely.

Close the old terminal session and reopen it so the new script is actually loaded into memory.

If anyone is confused about whether they are on Groq, NVIDIA, linked-account mode, or local direct-key mode, use the built-in runtime inspector instead of asking the model to guess.

If you do not want to link a Google account, the CLI can still run with a local Groq or NVIDIA key. The key stays in ~/.zerothink/config.json on that machine, not in the ZeroThink Vault. If your shell already has GROQ_API_KEY or NVIDIA_API_KEY, the CLI can auto-detect that too.

Direct mode is intentionally lean: no account memory, no saved ZeroThink history, no ZeroThink Vault key import, no linked-account session, and no Pro-only account features. Local CLI browser tools like zerothink search, zerothink browse, /search-web, and /page still work, but they are local terminal research helpers, not website account memory or account web-search state.

Use folder mode, then ask normal questions like you would in chat.

Pick Groq for the fastest practical coding default with openai/gpt-oss-120b, or switch to NVIDIA for the heavier coding lane like Qwen 480B Coder. You can also ask Agent Zero for the curated coding shortlist directly.

Attach a file or zip instead of pasting a giant wall of code into the terminal. If the repo contains a guidance file like ZEROTHINK.md, CLAUDE.md, AGENTS.md, or .claude/CLAUDE.md, the CLI also folds that guidance into local prompts automatically.

The CLI now ships with a real first-run guide plus beginner help topics so people do not have to memorize the command tree first.

Once you are inside the live prompt, the same guidance is available as slash commands, and unique short prefixes like /he can resolve to the right command. Steering and stop controls are built in now too.

The safest recovery pattern is to check identity, access, provider, and direct-key readiness before assuming the model is broken.

Run a live web search from the terminal, keep the numbered results in the current session, and reopen them later without leaving the CLI flow.

Open a saved result into the local browser state and preview the cleaned page text right inside the terminal.

Once a page matters, attach it into the current session so the next prompt can reason over that page alongside files and repo context.

You can fetch a live URL locally, search for a result, open a page into the browser state, and attach that page to the current session as prompt context.

For dynamic pages or real browser automation, ZeroThink CLI can register the official Playwright MCP server locally and drive it through a first-class wrapper instead of raw tool names. The built-in profile now uses headless isolated mode so separate runs do not fight over the same browser profile.

Drop a ZEROTHINK.commands.json file into a repo to add reusable prompt commands or machine-mode shell commands without editing the CLI itself.

Keep a clean step list inside the live session, checkpoint it with the rest of the conversation, and reopen the exact same working state later.

There is now a dedicated review wrapper for code-review style output instead of relying only on free-form prompting.

Use autopilot when you want the CLI to drive a multi-step task by itself instead of stopping after one reply. The controller chooses one next action at a time, saves state locally, and can finish with a proper summary.

Create named workers like zero1 and rerun them later. This is useful when you want a durable saved task with logs rather than a one-shot run.

The controller can spawn child workers like zero2 and zero3 for side tasks when that helps the main job. Those child runs are logged back into the parent agent state.

Agent mode can inspect local files, run shell commands, read and edit remote files on saved server profiles, search the web, and now call Playwright/MCP tools directly when machine mode allows it.

Read a file, search for symbols or TODOs, or scan a whole project tree before you ask for a refactor.

Create files, append notes, or do basic replacements without leaving the terminal.

Package a repo slice into a zip and attach that to a prompt when the project is too big to paste manually.

Save the host, user, port, and working directory so you do not have to retype them for every session. You can also use a saved password or an environment variable for password-based SSH when keys are not available.

Useful for status checks, logs, permissions, and quick repair work. Machine mode only.

Move zips, logs, build artifacts, or backups between your local machine and a saved target.

The CLI can now inspect and change text files on a saved server profile directly, which makes it much closer to a real operator tool instead of a chat wrapper.