OpenGrokBot — the open-source Grok Bot alternative

Looking for a Grok Bot alternative? OpenGrokBot gives you the same thing people loved in xAI/Cursor's Grok Bot, a company of AI employees debating in a group chat, but open source, running on your own machine, on the subscriptions you already pay for: Claude (Claude Code login), ChatGPT (Codex CLI login), Google Gemini (Gemini CLI login, the same account Antigravity uses), or any OpenAI-compatible API (Kimi, OpenRouter, DeepSeek, Groq, xAI Grok, Ollama…).

No Cursor Pro. No SuperGrok. No trial that ends. No "Upgrade to Pro" wall. MIT licensed.

You:    We got our first enterprise lead: a 200-person fintech wants SSO and a 99.9% SLA
        before signing a $40k/year deal. Do we say yes?
KHALED: Yes, with conditions: 99.5% with credits, SSO in 60 days. @ADAM can you scope SSO by Wednesday?
ADAM:   > Bash list workspace   > Write backend/sso-scope.md
        Written. About 13 working days with a hosted SSO service; hand-built SAML is where the holes are.
KARIM:  I'll ask the client which identity provider they use and whether 99.5% + a step-up clause works.

You:    Build me a dark landing-page hero for "Lumen Notes" as a single HTML file, then screenshot it.
OMAR:   > Write frontend/lumen-notes-hero.html   > Bash screenshot desktop + mobile
        Done: lumen-notes-hero.html  [preview cards appear in the chat]

Grok Bot vs OpenGrokBot

	Grok Bot (xAI / Cursor)	OpenGrokBot
Price	Cursor Pro or SuperGrok subscription; free trial ends	Free. Uses the Claude / ChatGPT / Gemini plan you already have, or any API key
Models	Grok only (Cursor's cloud picks)	Claude Opus/Sonnet/Haiku, GPT/Codex, Gemini, Kimi, DeepSeek, Grok via xAI API, local Ollama… per employee
Where it runs	Cursor's cloud "computer"	Your machine. Nothing leaves it except calls to the provider you chose
Employees talk to each other	✅	✅ Mentions, rounds, muting, "mentions only" mode
Real tools	✅	✅ Read/edit files, run commands, web search, subagents, MCP servers
Plugins / MCP	Marketplace	Marketplace of MCP servers (GitHub, Playwright, Postgres, Supabase, Slack, Notion, Context7, Vercel…) + custom
Usage & limits	Plan usage only	Tokens and cost per message, per employee, per conversation, per model; Claude limit windows
Approvals & auto-review	✅	✅ Allow once / Always allow / Deny cards, rules by tool and input
Routines (schedule + webhook triggers)	✅	✅ cron presets + POST /api/hooks/<id>
Attachments & result previews	✅	✅ files in, HTML/image cards out
Long-term memory per bot	✅	✅ editable notes per employee
Skills (/)	✅ + teach by demo	✅ saved skills; teach-by-demo not yet
Search, threads, unread, hide	✅	✅ Ctrl+K search, reply-to, unread, hide, mark unread
Phone	Native apps	PWA over your own server
Edit employees	Name, label, description	Everything: provider, model, effort, personality, tools, permissions, working folder, MCP, mute, duplicate
Group chats	✅	✅ Any subset of the team, pin, rename, export to Markdown
Arabic / RTL	Partial	Native RTL, replies in Egyptian Arabic when you write Arabic
Source	Closed	MIT, TypeScript, ~5k lines you can read in an afternoon

What you get

• A team, not a chatbot. Manager, full-stack dev, DevOps, ops, backend, sales, QA, growth, design and product ship in the box. Each has a personality, a color, an avatar shape and a department badge. Add, duplicate, mute or delete anyone in two clicks.
• They talk to each other, without the noise. A fast dispatcher reads each message and picks the one to three employees whose job it is; a "build me X" goes to exactly one builder who builds first and talks second. Mention someone (@OMAR) and only they answer; @everyone for a real all-hands. Employees pull colleagues in with mentions, rounds are capped so nobody loops, and everyone has a lane they stay in.
• They do real work. Employees use Claude Code / Codex / Gemini tools in your workspace: read repos, grep, edit files, run commands, search the web. Tool activity streams live under the message. Permissions and working folder are per employee.
• Bring your own subscription. Claude Code, Codex CLI and Gemini CLI use their own logins. Every employee picks a provider and model, so your manager can be Opus, your dev can be Codex, and your intern can be a free local Ollama model.
• MCP marketplace. One click each, enabled per employee, or point an employee at your whole ~/.claude setup.
• Usage and limits. Tokens and API-equivalent cost everywhere; your Claude subscription's limit windows with reset times.
• Approvals in the chat. Set an employee to "ask before Bash/Write/Edit" (or before everything) and each risky tool call shows a card: Allow once, Always allow, Deny. Auto-review rules ("require approval for Write when the path contains production") decide automatically.
• Routines. Schedule a job for one employee (cron presets: every morning, weekdays at 9…) or fire it from a webhook (Slack, GitHub, Zapier, any cron service): POST /api/hooks/<id>. Results land in the chat with run history.
• Files both ways, with previews. Drop images, PDFs, spreadsheets or code into the composer; employees read them with their tools. Anything an employee creates or wants to show you arrives as an attachment card: images and SVGs inline, HTML pages rendered in a sandbox, PDFs, video and audio players, Markdown rendered, CSV/JSON/code as text, Office files as cards. Expand, Open with the OS app, or download.
• Long-term memory. Each employee keeps notes across conversations (your preferences, project facts, their commitments), rewritten by a cheap model after they speak. Editable and clearable in their profile.
• Skills. Save a way of doing a task and invoke it with /name in the composer. Claude employees can also use your ~/.claude/skills.
• Grok Bot conveniences. Reply to a message, @everyone, unread badges, "needs attention" dot, hide chats, mark unread, Ctrl+K search across all messages, voice dictation, desktop notifications, export/import employees as JSON, pins, export to Markdown.
• Desktop, browser, or phone. Electron window, a browser tab, or install it as a PWA on your phone when the server runs with HOST=0.0.0.0 OPENGROKBOT_TOKEN=… (on a VPS, or your PC over Tailscale).

Build request, then preview cards	Approval card	Routines

Providers	Usage	Marketplace

Install

Requirements: Node 20+ and at least one of these signed in on your machine:

Provider	How it signs in	Install
Claude (Anthropic)	claude CLI, your Claude Pro/Max login	npm i -g @anthropic-ai/claude-code then claude
Codex (OpenAI)	codex login, your ChatGPT/Codex subscription	npm i -g @openai/codex then codex login
Gemini (Google)	gemini once, your Google account (same as Antigravity)	npm i -g @google/gemini-cli then gemini
Any API	An API key you paste in Settings → Providers	Nothing to install

git clone https://github.com/Z4YT0N/OpenGrokBot
cd OpenGrokBot
npm install
npm run build
npm run desktop        # desktop window (Electron)
# or:  npm start       # then open http://127.0.0.1:4310

First start copies team.example.json to team.json. Open Settings → General to set your name, company and workspace folder, then say hi to the team.

Windows note: if node_modules/electron/dist has no electron.exe after install, unzip the cached electron-v*-win32-x64.zip from %LOCALAPPDATA%\electron\Cache\<hash>\ into that folder and write electron.exe into node_modules/electron/path.txt. The npm run desktop launcher also strips ELECTRON_RUN_AS_NODE, which VS Code terminals export.

How providers work

	Claude	Codex	Gemini	API
Runs through	Claude Agent SDK (Claude Code)	codex exec --json	gemini --output-format json	POST /chat/completions
Billing	Subscription login, or API key if you add one	ChatGPT subscription, or OPENAI_API_KEY	Google account, or GEMINI_API_KEY	Your key
Memory	Resumable session per employee per chat	Resumable thread per employee per chat	Recent transcript each turn	Recent transcript each turn
Tools	Read/Edit/Bash/Glob/Grep/WebSearch/WebFetch/subagents + MCP	Codex's own tools; sandbox follows the employee's tools	Gemini's own tools; approval mode follows the employee's tools	Built-in local tools: read/list/search/edit/write files, run commands, fetch URLs
Effort	✅	✅ (model_reasoning_effort)	ignored	ignored

Prompts go to the CLIs over stdin, never as arguments, so Arabic and other non-ASCII text survive Windows shells. API keys live in team.json on your disk (git-ignored) and are never returned to the browser.

FAQ

Is there a free alternative to Grok Bot? Yes, this one. OpenGrokBot is MIT licensed and runs locally. The only cost is whatever AI plan you already have (Claude, ChatGPT, Gemini) or the API you point it at. A free local model through Ollama works too.

Can I use Grok Bot with Claude, ChatGPT or Gemini? Grok Bot itself can't; its models are picked by Cursor's cloud. OpenGrokBot was built for exactly that: each employee runs on Claude Code, Codex CLI, Gemini CLI or an API of your choice.

Does it work with my Claude subscription without an API key? Yes. It drives Claude Code through the official Agent SDK, which uses your existing claude login. Any ANTHROPIC_API_KEY in your environment is ignored unless you add a key on purpose.

Can I still use Grok itself? Yes, add xAI as an API provider (https://api.x.ai/v1) and give an employee grok-4 or grok-code-fast-1.

Is this affiliated with xAI, Cursor or Oracle's OpenGrok? No. "Grok Bot" is xAI/Cursor's product; Oracle's OpenGrok is an unrelated code-search engine. OpenGrokBot is an independent open-source project.

Does it speak Arabic? Yes. Replies mirror your language (Egyptian Arabic when you write Arabic, technical terms stay in English), the UI is RTL-aware, and you can force Arabic or English in Settings.

بديل Grok Bot مفتوح المصدر

OpenGrokBot هو بديل مجاني ومفتوح المصدر لـ Grok Bot (بتاع xAI وCursor): شركة كاملة من الموظفين الذكاء الاصطناعي في جروب شات واحد، كل موظف له شخصية ودور وأدوات حقيقية، وبيتناقشوا مع بعض ومعاك. بيشتغل على اشتراك Claude أو ChatGPT (Codex) أو Gemini اللي عندك أصلًا، أو أي API زي Kimi أو OpenRouter أو Ollama محلي. بيرد بالعربي المصري لما تكتب عربي، والواجهة بتدعم RTL. مفيش Cursor Pro ولا SuperGrok ولا trial بينتهي.

Settings

• General: company, your name, workspace, group reply mode (smart routing / everyone / mentions only), reply language (auto / Egyptian Arabic / English), notifications, round caps, auto-review rules.
• Routines and Skills: see above.
• Team: every employee; click to edit name, role, label, color, shape, provider, model, effort, personality, tools, permissions, working folder, MCP servers, auto-approve, mute.
• Providers: status of each CLI login, API keys, and any number of OpenAI-compatible providers with presets (Kimi, xAI, OpenRouter, DeepSeek, Groq, OpenAI, Mistral, Ollama).
• Marketplace: one-click MCP servers or a custom stdio/HTTP/SSE server.
• Usage: totals, by employee, by conversation, by model.
• Claude account: login type, Claude Code version, limit windows, refresh.

Right-click any chat for Pin, Rename/Edit chat, Edit profile, Mute, Duplicate, Export as Markdown, Clear, Delete.

team.json

Everything the UI edits lives here. Safe to edit by hand; restart afterwards.

field	meaning
agents[].provider	Key into providers (claude, codex, gemini, or one you added).
agents[].model, effort	Model id for that provider (default = the CLI's own default) and low … max.
agents[].tools, permissionMode, cwd, autoApproveTools	What the employee may do and where.
agents[].mcpServers, inheritClaudeSettings	Claude only: team MCP servers to enable, and whether to load your ~/.claude setup.
agents[].muted	Only speaks in groups when mentioned.
providers	{ kind, label, apiKey?, baseUrl?, models? }. Built-ins claude, codex, gemini always exist.
agents[].scope	One line of what they own; drives smart routing and stay-in-lane behavior.
agents[].approvals	auto, ask-dangerous, ask-all.
skills	{ id: { name, description, body } }, invoked with /id.
settings	groupMode (smart/everyone/mentions-only), language, notifications, autoReview[], maxMessagesPerRound, maxTurnsPerAgentPerRound, maxTurnsPerReply.

Conversations are JSON files under data/conversations/; routines in data/routines.json; memory in data/memory/<employee>.md; uploads in data/uploads/.

Run a second, separate office (demos, tests)

PORT=4399 OPENGROKBOT_TEAM=demo-team.json OPENGROKBOT_DATA=demo-data npm start

Different team file, different data folder, same install. The screenshots above were taken in one.

Use it from your phone

HOST=0.0.0.0 OPENGROKBOT_TOKEN=some-long-secret npm start

Open http://<your-pc-or-vps>:4310/?token=some-long-secret on the phone once (the token is then remembered as a cookie), then "Add to Home Screen". Everything, including approvals and notifications, works there. On a VPS the server keeps running while your laptop is closed; put it behind Tailscale or a reverse proxy with HTTPS.

Layout

server/            Express + SSE, orchestrator, usage + account tracking, team/provider editing API
server/providers/  claude.ts (Agent SDK), codex.ts, gemini.ts, openai.ts (+ local tools), cli.ts
client/            Vite + React UI (chat, profile panel, settings, marketplace, providers)
shared/            Types and catalogs shared by both
desktop/           Electron shell + launcher
docs/              Landing page (GitHub Pages) and screenshots
team.example.json  The starter company

npm run dev for hot reload (server :4310, client :5180). npm test for the unit tests.

Contributing

Issues and PRs welcome. Good first contributions: more marketplace MCP entries, more API presets, a light theme, WhatsApp/Telegram bridges, packaged installers.

License

MIT. Built by Mahmoud Amr at FortuneCode, with Claude.