CLI Assistant

Raycast AI on the Claude or ChatGPT plan you already pay for.

Claude and ChatGPT plans include a coding CLI: Claude Code and Codex. This Raycast extension runs your requests through whichever one is installed and signed in on your Mac, so you get the everyday Raycast AI features without a Raycast Pro subscription, an API key or per-token billing.

• Ask AI from root search, and Chat with follow-ups and history
• Fix Spelling and Grammar, Improve Writing, Make Shorter: rewrite the selection in place
• Translate as you type, or Translate and Replace the selection with one hotkey
• Summarize and Explain the selection
• Ask About Finder Selection: questions about files and folders (PDFs, logs, code), read-only
• Optional web search, reasoning effort and tools (read files, run shell commands after your approval), switchable mid-chat
• Any model your plan offers (Haiku, Sonnet, Opus, GPT-5.6…), switchable mid-chat, even between Claude and Codex

How it stays fast and cheap: the CLIs are coding agents, so the extension turns off their tools, MCP servers, settings files and agent prompts. That brings Claude Code's first answer from ~9 s to ~3 s and Codex's prompt from ~17.6k to ~4.5k tokens. Details in How it works.

Not affiliated with, built by, or endorsed by Anthropic, OpenAI or Raycast. It runs the official, unmodified CLIs; sign-in stays with them.

Install

Paste into a terminal:

curl -fsSL https://raw.githubusercontent.com/umitaltintas/raycast-cli-assistant/main/install.sh | bash

No Node, npm or git needed. The script downloads the prebuilt extension (~50 KB) from the latest release, puts it into Raycast's extensions folder and registers it with Raycast. It also checks which of Claude Code and Codex are installed and signed in. Run the same command again to update; chats and settings are kept. To uninstall, press ⌘K on any of its commands in Raycast → Uninstall Extension.

Then open the extension preferences and set your Primary Language.

Extensions installed outside the Store show up in Raycast as development extensions and may require being signed in to a Raycast account. The script sends the same registration notification as ray develop (raycast://cli/…). That is not a documented Raycast interface, so if it ever changes, installing from source below always works.

git clone https://github.com/umitaltintas/raycast-cli-assistant.git
cd raycast-cli-assistant
npm install
npm run dev   # once the extension shows up in Raycast you can stop it with Ctrl+C

Requires Node 22.22.2 or newer.

Spotlight (optional)

Add --spotlight to also get an Ask AI app that Spotlight finds:

curl -fsSL https://raw.githubusercontent.com/umitaltintas/raycast-cli-assistant/main/install.sh | bash -s -- --spotlight

Press ⌘Space, type "Ask AI", ask a question and the answer shows in a dialog with Copy and Ask Another. It uses the model last chosen in the extension and doesn't need Raycast to be open. Running the installer again keeps it up to date; to remove it, delete ~/Applications/Ask AI.app. (A Shortcut would be nicer, but signing one needs an iCloud account.)

Requirements

• Raycast v2 on macOS
• At least one of:

  • Claude Code, signed in with a paid Claude plan:

    curl -fsSL https://claude.ai/install.sh | bash
    claude   # then /login
    

  • Codex, signed in with a ChatGPT plan: brew install codex, then codex login. The ChatGPT app also ships a copy of Codex, which is found automatically.

Commands

Choosing a model: from the dropdown in the search bar in Chat and Translate, or with ⌘⇧R on an answer. Models are grouped by CLI (Claude Code, Codex). The list is not hardcoded: it comes from the installed CLIs and matches your plans, so new models show up on their own. The chosen model is shared by all commands and remembered; switching mid-chat keeps the context, even across Claude and Codex.

On an answer: ↵ paste to the active app, ⌘↵ continue in chat, ⌘⇧C copy, ⌘R regenerate, ⌘⇧R ask another model, ⌘. stop.

Options, in Chat and on answers, start from the preferences and are remembered per chat:

The navigation title shows what is on (e.g. My chat · Web · Tools). Translate and the rewrite commands always run without them, to stay fast.

Tips:

• Make Ask AI a fallback command in Raycast settings: anything you type in root search that matches no command becomes a question with ↵.
• Give your most used commands (Translate and Replace, Fix Spelling and Grammar) a hotkey.

Preferences

How it works

Every request runs claude -p in streaming (stream-json) mode. Claude Code is a coding agent, so everything a chat assistant doesn't need is turned off:

• tools (--tools ""), MCP servers (--strict-mcp-config), settings files (--setting-sources "") and slash commands;
• a short assistant system prompt (--system-prompt) instead of Claude Code's coding prompt.

In my measurements this brings the time to the first answer from ~9 s down to ~3 s. --bare would be faster but only works with API keys and ignores Claude account logins.

Codex runs as codex exec --json in a read-only sandbox; follow-ups use codex exec resume. Its coding-agent prompt is replaced with the assistant prompt (model_instructions_file), and the skills list, sandbox rules, apps, plugins, web search and other agent tools are turned off, which cuts a short request from ~17.6k to ~4.5k input tokens. Compared with Claude Code it is slower (~4.5 s for a short answer in my tests, almost all of it the round trip to OpenAI) and its answers arrive in one piece rather than streamed. When you switch between Claude and Codex mid-chat, the turns the other CLI answered are sent along as context, since each keeps its own history.

Options map to CLI flags. Claude: --tools WebSearch,WebFetch, Read,Glob,Grep with --add-dir, Bash, and --effort. Bash is never pre-approved: in -p mode a call that needs approval is denied and listed in permission_denials, so the extension shows that exact command and, once approved, runs it itself and sends the output as the next message. Codex: web_search="live", model_reasoning_effort, and its shell in the read-only sandbox (turned off entirely when not needed, ~640 fewer tokens); commands the sandbox would block are asked for in an approve block and handled the same way.

More details:

• Chats are Claude Code sessions (--session-id / --resume). Their files are kept in the extension's own folder so they don't mix with your real projects.
• Model list: Claude models come from the CLI's control protocol (the initialize request also behind the Agent SDK's supportedModels()), Codex models from codex debug models; no model calls, ~1 s, both in parallel. The result is cached and refreshed daily in the background; if the CLI is unreachable a fallback list of aliases is used.
• Translation direction: Auto mode is not left to the translation itself, which was unreliable. For longer text the extension detects the language with franc and sends one source and one target language; if the result comes back (almost) unchanged, the direction is reversed once. Short text is too ambiguous for franc ("benim adim ne" and "hello" were both misdetected), so a separate call (Haiku, or the chosen Codex model) only answers "primary or secondary?" while both directions are translated in parallel; the classified one is shown. That takes ~3 s, the same as a single translation.
• Wrong-language check: if a translation from a Claude model comes back in the wrong language (Haiku sometimes does this, e.g. German → Turkish), it is retried once with Sonnet and the result is marked. Codex requests always stay on Codex.
• Proxy: --setting-sources "" skips user settings, so HTTPS_PROXY / HTTP_PROXY / NO_PROXY from ~/.claude/settings.json are read by the extension and passed to the CLI.
• Raycast environment: Raycast extension processes have no USER / LOGNAME; without them Claude Code can't find its Keychain login and says Not logged in. The extension sets them.
• If ANTHROPIC_API_KEY is set, the CLI would use it instead of your subscription, so it is removed from the child process.

Terms

The extension does not collect, store or relay credentials. It only runs the unmodified Claude Code or Codex CLI that you installed, and sign-in happens through Anthropic's or OpenAI's own flow. Your usage is subject to their terms (Anthropic, OpenAI) and your plan's limits.

License

MIT