Grok Bot gives you named, persistent AI teammates sharing one always-on cloud Linux machine with a browser, a filesystem and a terminal. Most of the headline claims made for it are not differentiators: ChatGPT shipped Tasks in January 2025, its tasks now fire on connector activity rather than only a clock, and agent mode has had a cloud browser with a terminal since July 2025. What I could not find a counterpart for in a consumer assistant is the shape — a roster of named Bots with durable memory as the unit of work rather than one-off tasks. Teaching a skill by recording your browser is not itself new (Google's Project Mariner does it), but getting back an editable text skill rather than an opaque replay is unusual. The real constraint is cost: usage is metered with no spend cap, and driving a browser is the most expensive thing a Bot can do. Which is why you point it at a structured connector. Liftli is one MCP URL, and Grok support is live — I have verified grok.com and the Grok Bot desktop app end to end, including live tool calls.
Most of the Grok Bot coverage led with the same headline: AI agents that run on a schedule. I went looking for what was actually new, and a lot of it did not survive the search.
What Grok Bot actually is
You create named Bots — persistent teammates with their own instructions, memory and connectors. All of them share one always-on cloud Linux machine, one per account rather than one per Bot: a browser, a /workspace filesystem and a terminal that keep running when your laptop is shut.
Routines start work without you, on a schedule or on an event such as a new message in a watched Slack channel; cap of 50 per Bot. (Grok calls them Routines; confusingly, so does Claude Code.) Skills are reusable recipes. The Bot signs into tools as you through browser sessions rather than a credential vault, handing control back when it hits a password, 2FA or a CAPTCHA.
It launched on 11 August 2026 and is still labelled beta. Access comes with SuperGrok (from around $30) or a Cursor Pro, Pro+, Ultra or Teams plan; desktop app for macOS and Windows, iOS companion, no Linux or Android.
What is genuinely new, and what is not
Three claims get made for Grok Bot that do not hold up. Two ChatGPT features cover all three.
| How it gets described | What I actually found |
|---|---|
| The first AI agents that run on a schedule | ChatGPT shipped Tasks in January 2025 |
| It reacts to a Slack message, not just a clock | ChatGPT's tasks now fire on connector activity, not just a clock |
| A whole cloud computer — browser, terminal, files | ChatGPT agent mode has shipped one since July 2025, on the $20 plan |
That is a shorter list than the launch coverage suggested, and I would rather you hear it here than fifteen seconds after someone challenges you.
Two things I could not find a counterpart for:
The unit of work is a teammate, not a task. Everywhere else you schedule a job. Here you hire someone: a named Bot with durable memory that accumulates context about the account it works on across every run, and stages material in /workspace between them. That sounds like branding until you have four of them and notice you are managing a roster rather than a cron table. Custom GPTs are named and persistent too, but nothing happens until you open one; there is no run to remember. A Bot runs while you sleep and remembers what it found.
What that feels like is mundane and slightly strange. You open the app in the morning to a list of names, not a list of jobs. Scout found four things overnight; Writer is waiting on a decision. Neither stopped existing when its run finished. It is closer to reading your team's standup than checking whether a cron job exited zero.
Teach-by-demonstration, with a caveat. Record yourself doing a thing in a browser for ten minutes and get back a draft skill. Record-and-replay is old news in enterprise RPA, and Google's Project Mariner has shipped "teach and repeat" since 2025. What I have not seen elsewhere in a consumer assistant is the output: an editable text skill you can read, correct and hand to a Routine, rather than an opaque replay.
Neither is a moat. Both are useful, and they are enough to build on.
Where it burns money, and why that argues for a connector
The most common complaint about Grok Bot is not reliability. It is the meter. Usage sits in its own bucket with a weekly allowance, uncapped overage and no spend cap; users have reported burning a large share of a week's allowance on day one, one widely shared figure being 42%. Treat any single number as anecdote, but the pattern is consistent enough to plan around.
The expensive thing is the browser. Publishing a LinkedIn post by driving one is navigate, screenshot, click, type, screenshot, verify — dozens of metered steps, each a chance to break on a layout change. xAI says this itself: when a connector exists, use it, because it is the more reliable path. It is also the cheaper one. Liftli's distribute — Liftli is a content strategist your assistant calls over MCP — takes the text and returns the live post URL in one call.
Three more limits before you trust a Bot with anything that matters. Bots are not a security boundary (xAI says so plainly). Approvals are prose, not policy: no dry run, no undo, no audit log yet. And there are no compliance certifications — if you are in a regulated industry, this is not your quarter.
None of that is a reason to skip Grok Bot. It is a reason to be deliberate: connectors for the work, and a server-side gate on anything that becomes public.
Connecting Liftli to Grok
Liftli is a single remote MCP server, on the same URL every custom connector uses:
https://mcp.liftli.ai/mcp
On grok.com (verified on the live site): open Skills and Connectors, click New Connector, choose Custom — add your own custom connector. The form asks for exactly two things: a Name ("Liftli") and the Server URL above. No client ID, no secret, no token-auth dropdown — Grok handles registration itself. Click Add Connector and sign in to Liftli in the popup. Success is a connector card reading "All tools enabled". I ran this on a paid SuperGrok account; I could not confirm whether custom connectors are available on the free tier.
In the Grok Bot app there is no settings screen for this. The Plugins screen only browses the marketplace, so the only way to add a custom connector is to ask the Bot — which is itself a real signal about where Grok Bot is going: configuration is a conversation, not a form. Tell yours to use Liftli and it offers to add the connector itself; I watched it find the URL from Liftli's public docs, present an Add Liftli button, and connect after one authorize tap. If it does not offer, say it plainly: add https://mcp.liftli.ai/mcp as a custom connector. Then attach it in the composer with @Liftli.
Three things worth knowing. The two surfaces keep separate connector lists, so if you use both, add Liftli in each. Connectors are account-wide in Grok Bot, shared across all your Bots rather than scoped per Bot. And Grok Bot reportedly loads MCP tools lazily — a Routine that assumes they are present may find none loaded, so name @Liftli explicitly in the first step of any Skill or Routine.
On iPhone, Grok Bot can't add a connector itself yet — the sign-in completes and the card even reads Connected, but the app never attaches the token, a Grok Bot beta bug. Do the add once from the desktop app and the iPhone works too: connectors are account-wide within Grok Bot. grok.com in a browser is a different door — Liftli in Grok chat, not Grok Bot.
Two workflows worth setting up this week
Liftli decides what is worth saying, drafts in your voice, and holds a publish gate you control. A "seed" is one raw idea it thinks is worth a post.
1. The overnight extraction routine
Ghostwriting is roughly 20% writing and 80% extraction. Every tool in this category automates the writing and then sells you a manual ten-minute interview for the extraction. Unattended agents are not new. What I have not seen is one pointed at the founder's own working material with a strategist deciding what is worth keeping.
Create a Bot called Scout. Attach only the connectors you are comfortable with — Slack, Gmail, Calendar, Notion — plus @Liftli. Give it a nightly Routine, or trigger it on new messages in one channel.
Each run, Scout pulls the day's raw material into /workspace, then asks Liftli for the seed-extraction brief: what counts as a seed, what is noise, how to score it. It files each keeper with remember(...), storing the full transcript alongside the one-line idea, then calls channel(action='digest', items=[...]) to send one message per seed with a tappable "Add take" button. In the morning you tap one and talk for twenty seconds; that take binds to that seed, so when you later plan a post from it, the raw material and your twenty seconds load together.
The architectural point: Liftli never touches your inbox or your chat tools. Slack, Gmail and Calendar connect to Grok, not to Liftli. The digest is a receipt that nothing moved without you seeing it — a receipt and not an approval queue, because gating seeds before they flow just reinvents the ten-minute interview.
Two prerequisites, so you do not meet them at 2am. This is a paid-tier workflow: the seed-extraction brief is metered at three free runs, lifetime. The credit is keyed to the material, so a re-run on the same input does not charge twice — but a nightly routine sees new material every night, which is exactly three nights free. And the digest needs the push relay enabled plus a linked Telegram; without both, seeds still get saved, they just do not reach your phone.
2. The gate that cannot be skipped
The moment you tell someone an always-on AI can post for them, they picture the disaster. The answer is not reassurance, it is mechanism. Liftli's review gate is a server-side state machine, not an instruction:
ready_to_postreturns STOP unless this exact draft was run throughcritiqueand the executed review came back with a pass verdict. The gate wants the verdict submitted, not described.- Editing the draft after the review voids the review, and
distributerefuses text that never cleared the gate — unless you typed the words yourself. - A model rewrite clears human approval; a human edit keeps it. The agent cannot quietly swap the words after you blessed them.
- Approving through the agent or the Telegram card emits a telemetry event — a record of who approved and when, which Grok Bot does not yet provide.
Wrap that sequence in a Grok Skill and the Routine that runs at 3am inherits it. The limit, stated straight: this gate covers Liftli's publish path. If you tell the Bot to browser-drive LinkedIn directly, it routes around every check, so your Skill should forbid the browser path for publishing — a discipline rather than something Liftli can enforce. xAI's own use-case catalogue stops short of autonomous publishing and recommends approval boundaries for it. I agree: the correct terminal action for an unattended loop is hand the human a card.
One thing I would not automate yet is a comment sweep that posts by itself. Draft the comments, shortlist them, post them yourself — automated commenting under your own session is the fastest route to a restricted account.