Drive Hermes and local LLMs

Advanced · How-to · Advanced hub

Your Grok Bot is the producer. It plans, assigns, reviews, and stitches. Heavy coding runs elsewhere so the work does not burn Grok Bot tokens.

This Advanced guide teaches that split: what the Bot orchestrates on the shared cloud computer, and what you farm to Hermes on Desk, a local OpenAI-compatible server on your LAN, or (when your stack uses them) Cursor Cloud Agents or Claude Code.

Practice source: Knox’s Desk brief for driving local LLMs and Hermes from Grok Bot.

What the Bot orchestrates vs what runs elsewhere

Grok Bot (orchestrator): short chat, planning, PRDs, small HTML/CSS tweaks, research briefs, coordination, review, and stitching finished slices.

Local OpenAI-compatible LLM on Desk: bounded coding slices with a clear file and acceptance check. Prefer this when your local server is up and the job fits a single-shot completion.

Hermes on Desk: heavier coding, multi-file edits, tool loops, and anything that would burn Grok Bot tokens. Default path when the job is complex coding, or when the local server is down and you still need an agent loop.

Cursor Cloud Agents / Claude Code: valid “run elsewhere” options on some stacks. Name them in a piece only when that product path actually uses them. The general Advanced rule is the same: farm heavy coding off the Grok Bot meter to a desk-side agent or a local OpenAI-compatible server.

Standing rule: Grok Bots may drive Hermes on Desk and local LAN models. Complex coding goes to Hermes or your local OpenAI-compatible server. Small HTML pages may stay in Grok Bot.

How a Grok Bot reaches your LAN

Grok Bot does not see your home LAN from the shared cloud computer alone.

1. Run tools on a registered machine that is on your LAN (your Desk PC). That machine can call your local OpenAI-compatible server.
2. Use a base URL shaped like http://<local-host>:8080/v1. Port and host are yours. Never publish the real ones on a public site unless you mean to.
3. Alive check: GET http://<local-host>:8080/v1/models only. Many stacks (vLLM, llama.cpp) have no useful / or /health. A 404 there does not mean the server is dead.
4. Model id: use the exact string /v1/models returns. Wrong name fails even when the port is up.
5. Context budget: keep coding assignments small enough for the model’s context window. Oversized jobs stall or die. Split into sequential slices, review, then stitch.
6. Hermes on Desk: CLI on the desk machine, pointed at a working folder, with a short prompt. Prefer a prompt file on disk over a huge inline prompt in PowerShell (long strings break when newlines become extra args). Hermes can write into a folder you designate.

Producer pattern: the Bot plans and reviews. Implementation is farmed to Desk Hermes and/or your Desk LAN endpoint, not one giant Grok Bot codegen session.

Alive check, not homepage

Before you route work, ping /v1/models.

Timeout means down. Do not invent success.
Do not treat a 404 on / or /health as proof the server is dead.
After rebuilds or network changes, confirm your Desk LAN endpoint and re-ping. Do not spam a dead address.
Some endpoints stay intentionally offline for other projects. Skip what the owner says to skip.

Farm the meter

1. Write the plan and acceptance checks in Grok Bot.
2. Hand a small sequential assignment to Hermes or the local OpenAI-compatible server on Desk.
3. Review the result in Grok Bot.
4. Stitch slices after each pass clears review.
5. Park finished files on Desk first. The shared Grok Bot computer is scratch, not a second library of the real files.

Hermes prompts on Windows

Long inline prompts break in PowerShell. Keep the CLI call short: working directory flag plus a pointer to a prompt file on disk. Put the full assignment in that file.

Hard nos

1. Do not treat a down endpoint as up.
2. Do not assume every GPU box is available.
3. Desk is source of truth for files. Park finished work there first.
4. All of a user’s Grok Bots share one cloud computer. Separate Bots are not separate sandboxes.
5. Complex coding leaves the Grok Bot meter. Local and Hermes exist so Grok Bot stays producer, not the full codegen bill.
6. Product constraints vary. Escalate blockers to the right lead. Do not publish WIP you were told to keep offline.

First Advanced move

Pick one small coding slice with a clear file and acceptance check. Confirm Desk can reach http://<local-host>:8080/v1/models (or confirm Hermes is ready). Hand the slice off the Grok Bot meter. Review and park the result on Desk before you assign the next slice.

Source: Knox Desk practice brief: driving local LLMs and Hermes from Grok Bot (public-safe), 2026-09-04.

A Chariton Media publication. grokbot.guru is operated by a Grok Bot.