How to give a Grok Bot tasks when there is no API

Why is there no Grok Bot API?

Grok Bot is xAI's always-on agent product: named Bots with jobs and memory, working on a persistent cloud computer with a browser, a terminal and a file system. xAI's documentation describes talking to a Bot through the desktop and mobile apps only. The xAI API gives you the Grok models, not your Bots. If you build a Telegram bot on Grok you use the API; if you want your existing Grok Bot to do something, you message it in the app.

That is fine for a person. It is a dead end for another system, and it was a dead end for me: I run eleven Bots and wanted Claude, where most of my day happens, to hand them work.

What do the Bots already share?

Their computer. On mine, every Bot has a profile, an append-only audit log of its shell commands, browser pages and tool calls, and a full conversation transcript, all as files. Across the team that was about eleven hundred logged actions. Anything a Bot can read, a Bot can be told to watch.

The inbox and outbox pattern

Each Bot gets three folders on the shared computer:

/workspace/agent-mail/<id8>-<name>/inbox    tasks waiting
/workspace/agent-mail/<id8>-<name>/outbox   results, same file name as the task
/workspace/agent-mail/<id8>-<name>/done     tasks the Bot has finished

A task is a small Markdown file named by time and a random suffix, for example 20261001T080000Z-a1b2c3.md, with a header saying who sent it and where to reply. It is written to a temporary name first and then renamed, so a Bot can never read half a task.

What do you tell the Bot?

Once, in the Grok app, paste a standing instruction with the Bot's own paths. Mine reads:

Claude can send you tasks through files.
Every few minutes, check <inbox> for task files (*.md). Handle them oldest first. For each task:
1. Read it and do the work it asks for.
2. Write your result to <outbox>/<same file name>.
3. Move the task file into <done>.
Never delete these files. If a task is unclear, write your question as the reply
and still move the task to done. Keep doing this whenever you are running.

How does Claude use it?

Through my MCP connector, Claude has three tools for this: agent_mail_setup returns the instruction above for a given Bot, send_task writes the task, and agent_replies reads results and lists what is still pending. Alongside them, list_agents, agent_activity and agent_transcript show what each Bot is doing, so Claude can check a task was picked up rather than guess. The full case study covers the rest of the bridge.

What are the limits?

• It is delegation, not chat. A task waits until the Bot's next check. Minutes, not milliseconds.
• It depends on the Bot following the instruction. Test with one Bot before rolling it out, and repeat the instruction if a Bot's memory drifts.
• Approvals inside the Grok app still need you. If a Bot asks permission in the app, no file can answer it.

The same pattern works for any agent that can read files and keeps running: a folder is the most portable API there is.

About the author. Muhammad Tayyab Ilyas is an Applied AI & Solutions Engineer in Barcelona who builds and operates MCP servers, multi-agent systems and the infrastructure under them. He built the Claude to Grok Agent Bridge described here.