Grok Bot Reads Your Playbook Already

A new agent lands on your machine roughly every other week, and each one asks the same question: where do I put my skills this time? The honest answer for most of them has been "in another folder, in another format, kept in sync by hand."

Grok Bot, xAI's desktop agent, is a pleasant exception — and the grok target shipping in the AgentPlaybooks CLI today mostly just makes that fact explicit.

What we found

Grok Bot discovers skills from a fixed set of roots. Read out of the shipped application, that list is:

.cursor/skills/   .cursor/skills-cursor/   .agents/skills/
.claude/skills/   .codex/skills/           .claude/plugins/

.agents/skills/ — the portable store AgentPlaybooks has been writing all along — is in there, classified as a workspace skill source. And its system prompt loads AGENTS.md directly, telling the model that the file may carry environment details, code guidelines, "an overview of important rules or skills and when they should be used", and that it should always follow them.

Which means the interesting part of the integration was already done. If you have run agentplaybooks sync on a project, or pulled a playbook onto a new machine, Grok Bot sees those skills and those instructions with no bridge file, no copy step, and no format conversion.

The new target

agentplaybooks sync . --target=grok --apply

| Target | Skills | MCP servers | Instructions | |---|---|---|---| | grok — Grok Bot (xAI) | .agents/skills//SKILL.md | — (account MCP Box) | reads AGENTS.md natively |

You could already reach the same files through the antigravity target, since both read the portable store — but calling the Grok target "Antigravity" is the kind of small lie that costs someone an hour later. doctor and sync now detect the app by its ~/.grokbot profile and suggest the target by name.

The one thing a file cannot do

MCP servers. Grok Bot stores only an array of server ids in ~/.grokbot/settings.json:

{ "mcpBoxServers": [], "mcpCustomInstructionsByServerId": {}, "mcpDisabledToolsByServerId": {} }

The definitions behind those ids live in your account's MCP Box, not on disk. No project file provisions them, which means sync genuinely cannot deliver a playbook's MCP servers to this target.

What it does instead is say so:

[conflict:grok] mcp 'deploy': Grok Bot's MCP servers live in the account's
MCP Box, not in a project file. Add the playbook's own MCP endpoint
(POST /api/mcp/<guid>) to the Box once, and its tools reach every Grok Bot
session.

A playbook whose MCP servers quietly did not arrive looks exactly like a playbook that has none — and the person who finds out is the one debugging a missing tool at the wrong moment. So the CLI reports the gap rather than skipping it in silence. It reports it only when there is actually something it would have delivered; a playbook with no MCP servers says nothing.

The way around it is one entry, once. Add your playbook's own MCP endpoint — POST /api/mcp/ — to the Box, and everything the playbook holds arrives through that single server: skills, memory, canvas documents, and use_secret for credentials the agent should be able to use without ever reading. Per-server custom instructions and per-server disabled tools are local settings in Grok Bot, so you can still tune what that one server exposes.

Try it

npx agentplaybooks doctor .
npx agentplaybooks sync . --target=grok

Plan first; nothing is written until you add --apply. If the project already has skills in .claude/skills/ or .cursor/skills/, the plan will show them landing in .agents/skills/ — and that is the whole migration.