Grok Bot can write and run code. Its cloud Linux machine has a terminal, a filesystem, and a browser, so a Bot can clone a repository, install dependencies, edit files, and run a test suite. What it lacks is everything that makes agentic coding safe and reviewable: there is one shared computer per account rather than an isolated environment per task, no git worktree story, no diff or pull request surface, no model picker (xAI documents no default model for Bots and gives you no way to choose one), and no dry run, so a rehearsal performs real work. On the most product-realistic public benchmark, SWE-Marathon v1.1, Grok Build paired with Grok 4.6 scores 31.9% against Claude Code paired with Opus 5 at 50.0%. xAI’s own coding product is Grok Build, a CLI, and it is the right tool from xAI for repository work. The honest recommendation for coding is to run coding agents, Grok Build included, on hardware you control.
- Yes, technically. The documented cloud machine is a Linux VM with a browser, filesystem, and terminal, so a Bot can clone, edit, build, and test.
- No isolation between jobs. The docs state all your Bots share one computer, its files, its browser sessions, and its logins. There is no per-task sandbox and no worktree.
- No review furniture. No diff view, no staged hunks, no pull request flow, no plan gate. You read the transcript and trust it.
- No model picker, no default published. You cannot know which model wrote your code, and you cannot pick a cheaper one for cheap work.
- A rehearsal is not a rehearsal. There is no dry-run mode, so a test run pushes the branch, sends the message, and files the ticket for real.
- xAI already ships the coding tool. Grok Build is the CLI, it reads
AGENTS.md, it supports MCP and ACP, and it runs on a $30 SuperGrok plan rather than a Grok Bot tier.
The literal answer: yes, it has a terminal
Grok Bot’s documentation is unambiguous about the machine. From docs.x.ai/grok-bot/overview, retrieved 22 August 2026: Bots run "on a persistent cloud VM with a browser, filesystem, and terminal". A terminal on a Linux box with a filesystem is a development environment. There is nothing stopping a Bot from running git clone, installing a toolchain, editing a file, and running a test suite, and people did exactly that in launch week.
So if the question is whether the capability exists, it does, and any page telling you Grok Bot "cannot write code" is wrong. The interesting question is the one a working developer actually asks, which is whether it is a good place to do repository work. That answer is no, and the reasons are structural rather than a matter of the model being weak.
Where it fails for real repository work
Four gaps, in the order they will bite you.
| What a coding agent needs | What Grok Bot has | Consequence |
|---|---|---|
| An isolated checkout per task | One shared machine per account, shared files and logins | Two Bots on two branches share one filesystem. A half-finished refactor is visible to everything else on the account |
| A diff to read before anything lands | A chat transcript | Review means reading prose about what was changed rather than the change itself |
| A model you selected | Automatic routing, no picker, no documented default | You cannot attribute an output to a model, reproduce it, or route cheap work to a cheap model |
| A rehearsal that changes nothing | No dry-run mode | A test run performs the real action. Rehearse a push and you have pushed |
The isolation gap is the load-bearing one. The documentation states it plainly: "All of your Bots use the same persistent cloud computer. They share files, browser sessions, and app logins, which makes handoffs possible without repeating setup." And then, on the security question: each Bot gets its own screen "without getting separate security boundaries". For handing a research file to a writing Bot, that sharing is a feature. For running two agents on two branches of the same repository, it is the exact thing worktree isolation exists to prevent. There is no documented way to give a coding task its own filesystem.
The review gap is what makes the isolation gap expensive. A coding workbench answers "what did it actually change" with a diff, hunk by hunk, before you commit. Grok Bot answers it with the Bot’s own account of its work. eesel’s review notes that the audit view is described in the docs as coming, twice, and that today there are only per-Bot transcripts. Reading an agent’s description of its own edits is not review, and every experienced user of these tools learns that the hard way.
The undisclosed model router, and why it matters more for code
Three facts, and only the first is widely reported.
Grok 4.6 shipped on 12 August 2026, one day after Grok Bot, with a 500K token context window, a 1 February 2026 knowledge cutoff, and a new xhigh reasoning effort. Cursor’s own launch post frames it as optimised for "long-running agents and more ambitious interactive and visual work" and puts its published API rates at $2 per million input tokens and $6 per million output. It is live in the xAI API, in Grok Build, in Cursor, and in Grok Bot.
Second: xAI never states which model powers a Bot. Not on the launch post, not on the product page, not on any Grok Bot documentation page. Eleven days after launch, nobody has corrected that.
Third: there is no picker. Grok Bot routes automatically, and VentureBeat criticised the opacity directly, quoting a tester who found the routing "wasn’t great" in the early days.
For a research digest, an opaque router is a mild annoyance. For code it is a correctness problem. When a refactor comes back wrong you cannot tell whether you got the reasoning model or a faster one, you cannot re-run the same task against the same model to isolate the variable, and you cannot bill the cheap parts of your week to a cheap model. Every serious coding workflow eventually becomes a routing decision, and this product removes the decision.
The benchmark gap: SWE-Marathon
Benchmarks in this space are noisy, and the honest way to use them is to pick the one that most resembles the work and report the pairing explicitly. That benchmark is SWE-Marathon, because it scores a product plus a model rather than a model in a shared harness, which is what you actually buy.
| Pairing | Score | What it is measuring |
|---|---|---|
| Claude Code with Opus 5 | 50.0% | A dedicated coding agent with a coding-tuned frontier model |
| Grok Build with Grok 4.6 | 31.9% | xAI’s own coding CLI with its current flagship |
Two honest caveats. This is xAI’s coding product being measured, not Grok Bot, and Grok Bot is not on the board at all because nobody benchmarks it for code. And Grok 4.6 is not a weak model: on CursorBench v3.2 at xhigh effort it reports 70.8% against Claude Opus 5 at max effort on 70.0%, at roughly a third of the cost per task. The 18-point SWE-Marathon gap is about the agent scaffolding and the long-horizon task shape, not raw model quality.
Grok Bot vs Grok Build: use the right xAI product
If you are here because you want an xAI agent that works in a repository, xAI ships one and it is not Grok Bot. Grok Build is the CLI, announced 25 May 2026, with a documented install one-liner on macOS, Linux, and WSL, and a PowerShell equivalent on Windows. It reads AGENTS.md, supports plugins, hooks, skills, and MCP servers, runs subagents in worktrees, has a headless mode for CI, and speaks ACP so other tools can drive it.
The access path is also cheaper and simpler. xAI’s own announcement lists availability as SuperGrok and X Premium Plus subscribers, which is the $30 tier, not a Grok Bot tier. You do not need a Cursor account, you do not need Cursor Pro+ or Ultra, and you are not linking two accounts together permanently.
| Grok Bot | Grok Build | |
|---|---|---|
| Shape | Chat with named agents on a cloud VM | A CLI in your terminal |
| Where the code lives | xAI’s shared cloud machine | Your machine, your checkout |
| Isolation per task | None documented | Subagents in worktrees |
| Plan and diff review | Chat transcript only | Plan, review, approve with diffs |
| Model control | None. Automatic routing | Configurable, defaults to Grok 4.6 |
| MCP | Yes, via the plugins panel | Yes, plus hooks, skills, ACP |
| Access | Cursor account plus an eligible plan | SuperGrok or X Premium Plus, roughly $30 |
| Runs while you sleep | Yes | Only while a session is open |
The full disambiguation, including the third thing called a Grok CLI (an unaffiliated community project on GitHub with several thousand stars that has nothing to do with xAI), is in Grok Bot vs Grok Build and the Grok CLI guide.
What "always on" should mean for a coding agent
The strongest argument for Grok Bot in a developer’s hands is not the model. It is that the machine stays awake. The objection was stated on Hacker News by phoghed, on 20 August 2026, better than any marketing page has managed: "I very much don’t want to run it locally. I want the same one running somewhere else that I can interact with from all my devices. Look at something like Grok Bot. Nobody is going to run this locally. You can already self host almost anything, yet most people and businesses don’t."
That objection is correct as far as it goes, and it is worth answering rather than dodging. What it actually asks for is three properties: work that continues when your laptop closes, one place to watch it from, and control from a phone. Those properties do not require somebody else’s shared VM holding your logins. They require a host that stays up and a client that can reach it.
This is the design Continuum takes. You enroll a machine you already own, whether that is a Mac on your desk, a Linux box in a cupboard, or a small cloud instance you control. Coding agents run there under your own subscriptions, each Code session in its own git worktree, and you approve a plan, read a diff, or interrupt a run from an iPhone or a browser while the work keeps going on the host. The always-on property survives; the shared-machine trade does not have to come with it.
A practical recommendation
Sorted by what you are actually trying to do.
- You want repository work from xAI: install Grok Build. It is the product for this, it is cheaper, and it keeps your code on your machine.
- You want repository work from several labs: run Claude Code, Codex, Cursor, Grok Build, Gemini, and OpenCode side by side under a workbench that gives each session a worktree and shows you the diffs. That is the agentic development environment pattern.
- You want the non-code half of your week automated: Grok Bot is a reasonable fit, provided the shared machine holds no credential you would mind every Bot having. Read the security page before you sign anything in.
- You want a Bot to touch a repository anyway: give it a throwaway account, a fork rather than the main repository, and no deploy credentials. There is no dry run and no boundary between Bots, so scope the blast radius at the credential layer, because the product will not do it for you.
- You are choosing for a team: the blocking issues are not technical. There is no queryable organisation-wide audit view, no SSO story for Bot itself, and no published Grok Bot compliance claim found beyond terms that defer to Cursor. The enterprise page has the procurement checklist.
And one thing worth saying plainly, because it cuts against the easy conclusion: Grok Bot driving a browser on software with no API is a genuinely hard capability that most coding agents do not have. If that is your bottleneck, the right answer may well be to use both, with the Bot on the graphical work and a real coding agent on the repository.
Questions people ask
Can Grok Bot write code?
Yes. Its cloud machine is a Linux VM with a terminal, a filesystem, and a browser, so a Bot can clone a repository, install dependencies, write files, and run tests. What it lacks is everything that makes agentic coding reviewable: no isolated checkout per task, no diff surface, no pull request flow, no model picker, and no way to rehearse a change without performing it. The closest public evidence for the scaffolding gap is SWE-Marathon v1.1, where Claude Code with Opus 5 scores 50.0% against Grok Build with Grok 4.6 at 31.9%, and Grok Bot is not benchmarked for code at all. For repository work, xAI’s own coding product is Grok Build.
Is Grok Bot a coding agent?
Not in the sense the term is normally used. It is an always-on general agent that happens to have a terminal. xAI ships a separate coding agent, Grok Build, a CLI that reads AGENTS.md, runs subagents in git worktrees, supports MCP, hooks, and skills, and shows you a plan and diffs before applying changes. For repository work, Grok Build is the xAI product you want.
Can Grok Bot use GitHub?
It can, through a browser session or the terminal on the shared machine, and eesel’s review notes GitHub authentication failures on iOS during launch week. The bigger consideration is that a GitHub login typed once on that machine is available to every Bot on the account, present and future, because the documentation states Bots share browser sessions and app logins.
Does Grok Bot support git worktrees?
Nothing in the documentation describes worktree or per-task isolation. The documented model is the opposite: one persistent cloud computer per account, shared by every Bot, sharing files and sessions. You can create worktrees yourself from the Bot’s terminal, but they all live on the same shared filesystem, so they are not an isolation boundary in the way a coding workbench provides.
Which model does Grok Bot use for coding?
xAI does not say. Grok 4.6 is live in Grok Bot and is the plausible engine, but no default is documented and there is no model picker, so you cannot know which model wrote a given change or reproduce it against the same model later. For code specifically, that removes attribution, reproducibility, and cost routing all at once.
Is Grok Bot better than Claude Code for coding?
No, and the closest public evidence is SWE-Marathon v1.1, where Claude Code paired with Opus 5 scores 50.0% against Grok Build paired with Grok 4.6 at 31.9%. That measures xAI’s coding product rather than Grok Bot, which nobody benchmarks for code at all. Grok 4.6 itself is competitive per turn and cheaper; the gap is in the agent scaffolding on long multi-step tasks.
Can Grok Bot open a pull request?
It can perform the actions that produce one, by driving GitHub in the browser or running git and the gh CLI in its terminal. There is no pull request surface inside Grok Bot: no diff view, no staged hunks, no review pane. You review by reading the Bot’s own account of what it did, then checking the result in GitHub yourself.
What should I use instead of Grok Bot for coding?
Grok Build if you want xAI specifically, at roughly $30 on SuperGrok rather than a Grok Bot tier. If you want several labs’ agents at once with worktree isolation, diffs, live quota gauges, and phone control while the work runs on your own hardware, that is what Continuum does, and the app is free on Mac, iPhone, web, Windows, and Linux.
Sources
Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.
- Grok Bot overview, docs.x.ai the shared cloud computer, per-Bot screens, memory, routines, connectors and MCP, computer use, group chats
- Grok Bot get started, docs.x.ai eligible plans, Cursor sign-in, platform downloads, the Linux statement, Settings paths
- xAI: the Grok Build CLI xAI’s own coding agent: install, access tier, AGENTS.md, plugins, hooks, skills, MCP, ACP
- Cursor: Grok 4.6 Cursor’s own framing of what Grok 4.6 is optimised for, and its published rates
- Grok 4.6 release coverage context window, knowledge cutoff, published API rates, benchmark placement
- eesel AI: Grok Bot review metering behaviour, compliance gaps, no dry run, audit view described as coming
- VentureBeat on Grok Bot SpaceXAI branding, tier prices, criticism of automatic model routing
- Hacker News: "I very much don’t want to run it locally" phoghed’s comment, 20 August 2026, the clearest statement of the objection to own-hardware agents
- Hacker News: the Grok Bot launch thread launch-week reception, the trust and prompt-injection objections, the Linux download report
- Continuum pricing Free, Plus, Max 100, Max 200, Ultra, and the Team seat price