1. The tools in the box
Out of the box, Grok Build can read and edit files, search the project, run shell commands, and search the web. That set is already enough to change a repository. You do not need a plugin to get real work done, and you should not add one on the first day.
Every extra tool is more the model can do without you in the loop for that step. Add one when a repeated job needs it. Not because a list of integrations looked impressive.
2. A connected server adds tools you did not write
An MCP server is a program that offers tools over a standard protocol. A GitHub server might open issues. A database server might run queries. You configure servers in your Grok config, under mcp_servers. Once a server is enabled, its tools sit beside the built-in ones, and the model can call them.
Treat that program the way you treat software you installed. Who published it. What it can reach. Whether it needs a credential. Read the description before you enable it, and turn off the ones you are not using. A database tool pointed at a production database can read production. The protocol does not make that safe. Your configuration does, or does not.
Do not put a live secret in a note, a skill, or a handbook. If a server needs a key, keep the key in the place your team already uses for keys, and keep the sample config free of real ones.
3. Ask is the default, and it is the right default
By default, Grok asks before it edits a file and before it runs a command that changes something. Some read-only commands are allowed through. Reading a file does not stop for a prompt. A few commands that only create an empty file or an empty folder are let through too. A protected path still asks. When it does ask, you see the request and you approve it, or you do not.
That pause is the product working. The first week, leave it on. You learn what the loop tries to do. Approving a command you do not understand is how a session goes somewhere you did not intend. Say no, and ask what the command is for. “What does this change, in one sentence?” is a complete reply. If the answer is unclear, the command waits. A connected database tool gets one more question: which database, and is it a copy you can throw away?
4. Always-approve skips the questions. It does not skip the rules.
/always-approve lets ordinary tool calls run without a prompt. There is a launch flag for the same mode. Use it when you already trust the repository and you are watching the session. It is a poor setting on a project you just cloned, and a poor setting when a connected server can reach systems outside the repo.
Always-approve is not “no limits.” A deny rule can still block a command. A hook can still block a command. Those are checks you or your team set up. They are not a reason to stop reading. A command that was allowed can still be the wrong command.
5. The sandbox is off until you turn it on
The sandbox is an operating-system limit on what the session can touch. It is off by default. grok --sandbox workspace is the profile aimed at ordinary development: commands can write in the project, in Grok’s own folder, and in temporary directories. They can still read the rest of the machine, and they can still use the network. It is a boundary on where a command can write. It is not a review. A bad edit inside the project is still a bad edit.
A hook is a separate control. A hook that runs before a tool can refuse a command, which is how a team blocks something destructive even for people who have always-approve on. You do not need a hook to start. You need the ask default, a short list of servers you actually use, and the habit on the next page: make the change prove itself.