Grok is the name of xAI’s AI models and AI assistant. You can use Grok to answer questions, summarise text, research information and write code, for example. The model behind the assistant is a Large Language Model: software that generates a response based on your instructions and the available context.
For tasks on your own computer or server, there is also Grok Build. This coding agent combines a language model with tools for reading and editing files, running terminal commands and more. This means the agent does more than provide explanations: it can also carry out steps to complete a task.
On a Grok-VPS Grok Build runs as a local agent. By default, the AI model it uses runs externally. Installing the agent therefore does not mean that the Grok model is also hosted on your server.
Who develops Grok?
Grok was developed by xAI. Alongside its models and chatbot, the xAI ecosystem offers development tools such as Grok Build and an API. An API allows software to exchange requests and responses with a service without having to use a chat interface manually.
The name Grok therefore appears across several products. For server use, the distinction between the model and the agent is particularly important.
What is the difference between Grok and Grok Build?
| Component | What does it do? | Where is it used? |
|---|---|---|
| Grok models | Process prompts and context and generate responses or tool calls. | As the model behind an application or via an API. |
| Grok chatbot | Provides an interface for conversations, research and other AI features. | Via the website or app; available features depend on your account. |
| Grok Build | Carries out development tasks using a model and tools for your working environment. | Including in the terminal on your own computer or server. |
Grok Build and the Grok models are separate components. The agent works with xAI models but also supports custom model endpoints. Using another provider or a self-hosted model changes the intelligence powering the agent, not the role of the agent software.
Key features of Grok Build
- The agent carries out multiple steps, such as examining project files, proposing a change and running tests.
- Tools provide access to files, a terminal and additional services, subject to the configured permissions.
- An interactive terminal interface displays your conversation, tool actions and approval requests.
- Saved sessions allow you to resume an earlier conversation.
- Project instructions and skills define conventions and reusable workflows.
- The Model Context Protocol (MCP) lets you connect external tools.
- Alongside interactive use, Grok Build supports one-off tasks launched from scripts and other integrations.
How does Grok Build work?
Grok Build connects your instructions, the language model and the tools in your working environment. The model proposes actions. The software checks them against the configured access rules, carries out permitted tool actions and returns the results to the model.
For example, if you ask the agent to investigate an error in an application, a possible workflow looks like this:
- You describe the error and the desired outcome.
- The agent reads the relevant code and project instructions.
- The model analyses the information and proposes next steps.
- After obtaining the required approval, the agent might run a test or modify a file.
- The model reviews the tool results and the agent reports the changes.
- You check the outcome before putting the change into use.
Quality depends on the model, your instructions and the available context. Completing an agent task does not guarantee that the code is correct or secure. Always check AI-generated output for accuracy.
Which AI models does Grok Build use?
By default, Grok Build uses models from xAI’s range. A custom endpoint also allows you to connect other supported model services. An endpoint is the API address to which the agent sends its model requests.
For a self-hosted model, one option is Ollama The model must be suitable for the task and support the required features, such as tool calling: requesting tool actions in a structured format. Not all models support the same features or deliver the same results.
If you use an external model, your server does not need to perform the model computations. Its capacity requirements then mainly depend on local builds, tests and other project tasks. Running the model locally as well requires additional memory and processing power. Read more about hosting an LLM locally with Ollama.
Project instructions, skills and plugins
Project instructions describe how the agent should work within a project. In an AGENTS.md file, for example, you can specify coding style, test commands and rules about which changes the agent may make and when.
A skill contains reusable instructions for a task, potentially supplemented by scripts and reference files. Plugins bundle additional functionality, such as skills, tools or hooks. Hooks are scripts that run when particular events occur.
These instructions change how the agent works, but do not retrain the underlying language model. A skill or instruction file is not a technical security boundary either. Actual access depends on user permissions, approvals and sandbox settings.
Sessions and memory
Grok Build saves conversations and associated session data so you can return to an existing task later. This is different from a new conversation automatically knowing everything from earlier sessions.
There are also settings for memory across sessions. Availability and behaviour depend on your version and configuration. Treat saved conversations and memory as project data: they may contain code, tool results or other information that is not intended for everyone.
The Grok chatbot’s memory features are separate. A preference saved in an online Grok chat does not automatically become a project instruction for Grok Build on your server.
MCP and external tools
Through MCP, an external application makes functions available to the agent as tools. This allows you to connect Grok Build to a Git platform, database, internal API or infrastructure service, for example.
The available capabilities depend on the MCP server and the permissions granted. A read-only database connection has a different impact from an integration that changes data or manages infrastructure. The content and results of tool actions may also be sent to the selected model.
When should you use Grok?
The Grok chatbot is suitable for asking questions, writing, summarising and research. Grok Build is primarily designed for development work in which an agent actually works with files and tools.
Examples of Grok Build tasks include:
- explaining a codebase and locating relevant files
- investigating a bug and proposing a targeted change
- writing and running tests
- updating documentation based on project files
- analysing log files and configurations
- running recurring development tasks from scripts
For a text-only question, an agent with terminal and file access is usually unnecessary. Choose the interface and permissions that suit your task.
Using Grok Build on a VPS
With Grok Build on a VPS, the server provides a persistent development environment containing your project files and tools. You access the terminal remotely via SSH.
If you have a Grok VPS with the agent pre-installed, you can start with account or API configuration. On another Linux server, you first need to install the agent. Our guide “Getting started with Grok” explains how to install Grok Build, complete the onboarding process, optionally use your own model and work with the main administration commands.
Having an always-available server does not mean every task can run autonomously without limits. The agent may stop for required approvals, connection problems or usage limits. The selected model and connected services still determine what the agent can do.
Is Grok free?
The Grok chatbot offers free usage options and paid subscriptions. Features and limits vary by offering. API usage has its own access and billing terms; do not assume that a chat subscription also covers API usage.
The Grok Build source code is available under the Apache 2.0 licence. This does not automatically make model usage, external tools or server capacity free. A local model also requires hardware, storage and maintenance.
The agent’s licence also says nothing about the availability or licence of a specific Grok model. Check these separately if you want to use model weights yourself.
Privacy and security
With Grok Build, the agent runs on your own system, but data does not necessarily remain local. When using an external model, the agent sends the context used for a model request, potentially including code and tool results. Plugins, MCP servers and other services have their own data flows as well.
On Linux, the combination of user permissions, tool approvals and sandboxing determines which actions the agent can perform. A sandbox restricts technical access; an approval determines whether a tool action may begin. Grok Build does not have an active sandbox in its default configuration. Check and configure this before granting the agent access to sensitive projects.
Use a separate Linux user, grant only the necessary permissions and review changes before deploying them. Keep production data and credentials outside the agent’s accessible environment. See also the Grok sandbox documentation.
Grok encompasses AI models and an assistant for conversations and other AI tasks. With Grok Build, you use a language model as part of an agent that carries out development work on your own computer or server, using the tools and permissions you provide.