# Grok Bot inside the Agent OS

Grok Bot is not a replacement for Claude Code, Hermes or OpenClaw. It's a different shape of tool, and it's strongest when it sits **beside** them rather than instead of them.

This file is about where each job should go, and how to stop Grok Bot's output floating in its own island.

---

## What each one is actually good at

| | Best at | Weak at |
|---|---|---|
| **Grok Bot** | Always-on admin. Bots with roles, routines, group handoffs. Reachable from your phone. Zero setup. | Deep code work. No model choice. Usage-capped. Its files live on its own cloud machine. |
| **Claude Code** | Heavy building, refactors, long reasoning on your real repo. Skills, hooks, MCP. | Not always-on. Lives at your desk. |
| **Hermes** | Profiles per model, cheap/free routing, scripted pipelines, terminal-native automation. | You build the plumbing yourself. |
| **OpenClaw** | Browser and computer-use work on your own machine, with your own sessions. | Runs where your machine runs. |
| **Free Claude Code** | Volume work at near-zero cost. | Weaker on the hardest tasks. |

---

## Where to route what

**Send to Grok Bot:**
- Anything that must happen while you're asleep or away from your desk
- Inbox triage, calendar planning, invoice drafting, meeting-to-action
- Multi-step jobs across web apps with no clean API (teach-a-task shines here)
- Anything you want to fire off from your phone
- Trigger-driven work — a Slack message wakes a bot

**Keep in Claude Code / Hermes / OpenClaw:**
- Anything touching your codebase
- Long research with a big context you control
- Work where you need to pick the model (Grok Bot won't let you)
- Anything you'd rather not hand a third-party cloud machine a login for
- High-volume work where per-task cost matters — route that to free local models instead

**Rule of thumb:** if the job is *"be somewhere, watching, and act"* → Grok Bot. If the job is *"think hard about this thing in front of me"* → your desk stack.

---

## Feeding the shared Obsidian vault

The problem with adding a fifth tool is that its output lands somewhere the other four can't see. Fix that on day one.

**The pattern:** Grok Bot writes summaries out to a place the vault ingests, and the vault stays the source of truth.

**Three routes, easiest first:**

1. **Email to yourself.** Have a bot draft its daily/weekly summary to a dedicated address. Your existing vault pipeline files it. Zero new plumbing, works today.
2. **Board as the bridge.** Bots already write to your project board. If the vault already reads that board, you're done — the Grok Bot output arrives through a pipe you've already built.
3. **Zapier MCP → your notes tool.** Point a routine at Notion, Drive, or whatever your vault syncs from. Slower to set up, most direct.

**What to send, and what not to.** Send the *conclusions*: the decision, the number, the summary, the link. Don't try to mirror raw chat logs into the vault — that's noise, and it's the fastest way to make the vault useless.

Suggested vault landing spots:
- Daily briefs → `01 Daily/`
- Decisions a bot surfaced → `05 Memories/Decisions Log.md`
- Client/project state → the relevant `02 Projects/` note
- Anything reusable → promote to `04 Resources/`

---

## Keeping context in sync the other way

Grok Bot bots start blank about your business. Two ways to fix that:

1. **Put the context in the role prompt.** Who your clients are, what you sell, how you write. Boring but effective, and it's the only thing that survives a fresh bot.
2. **Give it read access to the same notes.** If your vault syncs to a connected tool (Drive, Notion), let the bot read it. Now both halves of your stack share one brain.

The second is what makes it feel like one system instead of five tools.

---

## A sane division of the week

| Time | Runs where |
|---|---|
| Overnight + early morning | Grok Bot routines — brief, inbox, competitor scan |
| Your working hours | Claude Code / Hermes at the desk on the hard thing |
| Triggered, any time | Grok Bot — Slack, Git, Linear events |
| Batch / volume jobs | Free local models via your own stack |
| End of day | Grok Bot meeting sweep → board → vault |

---

## The honest reason technical people buy it anyway

It's another pool of usage. If you keep hitting weekly limits on your other tools, this is a separate tank of agent work at a flat price — with the bonus that it's reachable from your phone.

That's a legitimate reason to run it. Just don't let it become a fifth island. Wire its output into the vault in week one, or it will.

---

*Part of the Grok Bot AI Full Course — <https://agentos.guide/grok-bot-course>*
