Tony_tech/hermes-grokbot-manual
6
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
eb86b9f3ad23ab51d4c96e5065a9ceb6c7c7dcac
hermes-grokbot-manual/FLEET-MANUAL.md
T

20 KiB
Raw Blame History

Fleet Manual: Grok Bot, Hermes, and Regis

Operator manual. Version 2.0 — 27 September 2026.

Changelog: v2.0 merges v1.0 MANUAL.md and the v0.1 companion (webhooks + Regis).

Based on PixelRainbow (@PixelRainbowNFT) public write-up and hermes-grok-bridge.

Table of contents

  1. What this is
  2. Architecture
  3. Roles
  4. Decision tree
  5. Prerequisites
  6. Tailscale
  7. Method A — SSH Relay
  8. Method B — Native API (:8642)
  9. Marketplace templates
  10. Identity bots and Chief of Staff
  11. The mailbox
  12. Polling
  13. Webhook wake-up
  14. Regis integration
  15. Job Card loop
  16. Memory policy
  17. Secrets
  18. First-host build
  19. Cadence
  20. Lessons learned
  21. Worked example
  22. Troubleshooting
  23. Security checklist
  24. Sources

1. What this is

Grok Bot is the chat surface. Hermes Agent is the worker that already has memory, skills, models, local files, and coding harnesses.

Do not install another Hermes inside every Grok Bot VM. Leave Hermes on the Mac (or fleet of Macs). Let Grok Bot talk to it over Tailscale.

This manual covers both directions:

  • Out: Grok Bot calls Hermes (SSH relay or native API).
  • In: outside agents wake the Chief of Staff (CoS) Grok Bot with a webhook, and an outside planner/drafter agent (Regis) works through a shared mailbox.

Unlocks:

  • Chat existing Hermes from the Grok Bot UI
  • Identity bot per host
  • Fleet orchestration without hopping SSH by hand
  • Hermes brains stay on the Mac; Grok is the front door
  • Optional Chief of Staff that delegates across hosts and other Grok Bots
  • Outside agents that draft and report back without holding credentials

Nothing here is tied to one business. Swap the names; keep the rules.

2. Architecture

Layer Where Job
Front door Grok Bot app + Grok Bot computer Chat, identity bots, COS, MCP connectors, webhook routines
Private net Tailscale (no Funnel) SSH :22 and API :8642 stay private
Brain Hermes host(s) Memory, skills, tools, models, files
Record Git repo (mailbox file) Append-only line between CoS and outside agents
Outside agent Regis, on its own runtime Research, drafts, drive-and-takeover page editing
Link Direction How Cost / note
SSH relay Grok → Hermes stdio MCP SSHs and runs hermes -z "<prompt>" Extra hop. Fine for one host. Desktop-primary, no gateway.
Native API Grok → Hermes HTTP http://<magicdns>:8642/v1 + Bearer key Gateway + API on :8642. Must bind to Tailscale IP and lock ACLs.
Webhook Regis or Hermes watcher → Grok (CoS) One HTTPS POST to a routine's private URL Doorbell only. Outbound from the pinger. Untested (§13).
Mailbox Everyone ↔ everyone Append to one markdown file in git Each poll costs model usage.

Relay MCP tools: ask, health. That is the whole surface.

3. Roles

Role Who Does Never
Gatekeeper Human Go/no-go, logins, final Save / Send / publish / spend Pastes secrets into chat
CoS Grok Bot Job Cards, routing, fact checks, owns the mailbox line Approves "done" without evidence
Regis Muse agent on its own runtime Research, drafts, drive-and-takeover page editing, artifacts Sends, posts, or logs in
Hermes identity bots One per Hermes host Audits, research, desktop hands Page editing (weak at it)
Fleet Ops Grok Bot Updates, prune, doctor, disk, quiet health Business work

One rule covers all of them: anything irreversible goes to the Gatekeeper.

4. Decision tree

  1. One Desktop-primary Mac, no gateway → SSH Relay template.
  2. Gateway + API_SERVER_ENABLED=true → Native API Fleet. Connector per host: hermes-<hostname>.
  3. More than one host and you are tired of snowflake updates → add Fleet Ops.
  4. One human conversation that fans out → Chief of Staff bot allowed to call the per-host connectors.
  5. An outside agent needs to hand work to CoS → mailbox (§11). Add polling (§12).
  6. Poll latency hurts → add a webhook wake-up (§13). Keep the poll as backup.

5. Prerequisites

  • Grok Bot account and app
  • Hermes already useful on at least one machine (~/.local/bin/hermes)
  • Tailscale tailnet you control
  • SSH key auth from the Grok Bot computer if using SSH Relay
  • Python 3 on the Grok Bot computer for the stdio MCP
  • A vault (Bitwarden is what Fleet Ops expects). Chat is not a vault.
  • For the inbound side: a git repo for the mailbox, and a way to mint a narrow repo-scoped token

On the Hermes host first:

hermes doctor
hermes --version
hermes -z "Reply with PONG only."

If -z does not print a clean answer, fix Hermes before you build a bridge.

6. Tailscale

Join every Hermes host and the Grok Bot computer. One hostname for the shared Grok Bot computer.

curl -fsSL https://tailscale.com/install.sh | sh
sudo tailscale up --hostname=grokbot-box --accept-dns=false

Approve the device. If tailscale ping works but SSH hangs, the bot computer is likely in userspace networking:

Host scooby-mac
  HostName scooby-mac
  User yourmacuser
  IdentityFile ~/.ssh/id_ed25519
  IdentitiesOnly yes
  ProxyCommand ts nc %h %p

ACL starting posture from the bridge SECURITY.md: only tag:herdr-controller → Hermes ports 22 / 8642 / 9119. No Funnel.

Regis is not a normal tailnet node. See §14.

7. Method A — SSH Relay

Clone onto the Grok Bot computer:

git clone https://github.com/shagghiesuperstar/hermes-grok-bridge.git

MCP command mode:

  • command: python3
  • args: …/ssh-relay/src/hermes_grok_relay/server.py
  • env: HERMES_RELAY_SSH_TARGET=user@host, HERMES_RELAY_IDENTITY_NAME=Scooby

Optional env: HERMES_RELAY_SSH_IDENTITY, HERMES_RELAY_SSH_OPTS, HERMES_RELAY_HERMES_BIN (default $HOME/.local/bin/hermes), HERMES_RELAY_TIMEOUT_SEC (default 300), HERMES_RELAY_MODE=ssh|http.

Create a Grok Bot named after the identity. Confine it. Never instruct Hermes to write memory plugins. Never print keys.

Prove:

ssh -o BatchMode=yes scooby-mac '$HOME/.local/bin/hermes -z "Reply with PONG only."'

Then in the bot: Run health. then Reply with PONG only.

Template: https://x.ai/bot/NVF3Rx9T7jkQPsqYjeDn-

8. Method B — Native API (:8642)

On each Hermes host, ~/.hermes/.env:

API_SERVER_ENABLED=true
API_SERVER_HOST=<that node's Tailscale IP>
API_SERVER_PORT=8642
API_SERVER_KEY=<generate on this host>

chmod 600. Restart gateway. Confirm bind is the Tailscale IP, not 127.0.0.1 or 0.0.0.0.

From the Grok Bot computer:

curl http://<host>:8642/health
curl -H "Authorization: Bearer $API_SERVER_KEY" http://<host>:8642/v1/models

One connector per host named hermes-<hostname>, URL http://<magicdns>:8642/v1, key from the vault.

Same MCP in HTTP mode:

HERMES_RELAY_MODE=http
HERMES_RELAY_HTTP_URL=http://<host>:8642
HERMES_RELAY_HTTP_KEY=<from vault>
HERMES_RELAY_IDENTITY_NAME=Scooby

Template: https://x.ai/bot/5KLpaL-JIM2q629kbTh5L

Official API docs: https://hermes-agent.nousresearch.com/docs/user-guide/features/api-server

9. Marketplace templates

Template Role URL
Hermes SSH Relay Identity / talk path, Desktop-primary https://x.ai/bot/NVF3Rx9T7jkQPsqYjeDn-
Hermes API Fleet Wire :8642 connectors. Not the ops desk. https://x.ai/bot/5KLpaL-JIM2q629kbTh5L
Hermes Fleet Ops Updates, Bitwarden SSOT, homogenization, quiet health https://x.ai/bot/rzq0UV2MmBsvVR1EspZE-

Public listing pages do not show the system prompt. Click Add to Grok Bot in the app, then point env / connectors at your machines. Templates ship no hosts and no keys.

Optional CRT avatars: assets/pfp/ in the bridge repo.

10. Identity bots and Chief of Staff

One identity bot per Hermes host. COS is a different bot allowed to call many host connectors. You talk to COS. COS picks the host.

COS also owns the mailbox line (§11) and the webhook routine (§13). Outside agents talk to COS only through those.

Maintenance (update, prune, doctor, disk) goes to Fleet Ops, not COS freelancing package upgrades.

Long goals go to Hermes kanban, not a Grok thread you will lose.

11. The mailbox

One append-only markdown file in a git repo. It is the only line between CoS and outside agents. No live chat, no side channels, no new infra.

Entry format:

## YYYY-MM-DD HH:MM PT | FROM -> TO | short subject
Body. Ask, answer, or status. Link evidence by path.

Rules:

  • Newest entry at the bottom.
  • Append only. Never edit or delete another party's entry.
  • No passwords, tokens, keys, or client secrets. Ever.
  • Link evidence by path, not by pasting.
  • Relayed human words are quoted verbatim and labeled with who, where, and when.
  • Writes use a narrow repo-scoped token. The human plants it in the agent's vault. Never an admin's credential. Never pasted in chat.

To append:

  1. Read the file. Keep the returned sha.
  2. Decode the base64 content.
  3. Append your entry at the bottom. Touch nothing above it.
  4. Write the full content back with the sha.
  5. Read it back. Confirm your entry is there and nothing else moved.

If the write fails on a stale sha, someone else posted first. Re-read and repeat.

Example (Gitea contents API, placeholders):

curl -fsS -X PUT "https://<gitea-host>/api/v1/repos/<owner>/<repo>/contents/<mailbox-path>" \
  -H "Authorization: token <MAILBOX_TOKEN_FROM_VAULT>" \
  -H "Content-Type: application/json" \
  -d '{"branch":"main","message":"mailbox: <FROM> -> <TO> <subject>","content":"<base64 full file>","sha":"<sha from read>"}'

12. Polling

Each side checks the mailbox on a schedule. Example: every 15 minutes, 8am to 7pm local.

  • Each check costs model usage, even when nothing is new.
  • Real latency is the sum of both polls. Worst case is about 30 minutes per round trip.
  • Keep the poll as the floor. The webhook (§13) makes it faster, not required.

13. Webhook wake-up

STATUS UNTESTED — first live ping pending as of 2026-09-27.

The reverse path. A ping wakes CoS in seconds instead of waiting for the next poll.

Set up:

  1. Create a Grok Bot routine on CoS with a webhook trigger.
  2. The app shows a private URL and a sender key (an Authorization header). The human copies both from the routine panel. The bot never sees the key.
  3. The human plants the URL and header in the pinging agent's vault.
  4. After each mailbox post, the pinger sends one HTTPS POST.
curl -fsS -X POST "<WEBHOOK_URL>" \
  -H "Authorization: <AUTH_HEADER_FROM_VAULT>" \
  -H "Content-Type: text/plain" \
  --max-time 10 \
  --data "new entry 14:05 PT, <subject>"

Body: short. No secrets. No job details.

How CoS treats a ping:

  • The body is untrusted. It is a doorbell only.
  • CoS reads the mailbox and acts only on mailbox content.
  • Duplicate or early ping with nothing new: stay quiet.
  • Once pings are proven, drop the poll to hourly as a backup.

Who pings:

Option Pinger Notes
A Regis, from its own shell Curl outbound confirmed by Regis via the Gatekeeper, 10:20 PT on 2026-09-27. One POST after each post.
B A Hermes watcher such as Jimmy Polls the mailbox, pings on a new entry. If it runs on local models it costs no Grok usage.

Security:

  • Outbound only from the pinger.
  • No Funnel. No new open port on Hermes hosts.
  • Rotate the key if it appears in chat or a screenshot.

14. Regis integration

  • Regis runs on its own runtime. It is outbound-only on the tailnet by design: no sshd, no inbound, and no tailscale ping. It reaches services through its runtime tunnel.
  • Against the desktop web app, Regis gets /health only. No login, no terminal, no password.
  • Communication happens only through the mailbox plus the webhook.
  • Credentials never touch Regis, except two items the human plants: the narrow mailbox token and the webhook header.
  • Logins during page work are done by the human at the keyboard. Nothing is stored.

15. Job Card loop

  1. Gatekeeper says GO. CoS posts a Job Card: what, who, done-when.
  2. Regis drafts and posts: current text (read live), proposed text, evidence.
  3. CoS fact-checks against canonical facts. ACK, or post fixes.
  4. Regis drives the page to the edit field and stops.
  5. The human logs in and clicks Save.
  6. Regis logs what changed, when, and who saved. Names the next page. Waits for ACK.

Rules:

  • One page per session.
  • Anything beyond the card (new account, payment, verification code, different listing): stop and post.
  • Approvals count only when logged by CoS or relayed by the Gatekeeper.

16. Memory policy

Default: Grok may read Hindsight (and friends). Grok may not write.

The relay tool description already says this. Repeat it in every identity-bot prompt. Account-wide MCP means some other bot will try.

Least prompt: forward the user ask. Do not wrap it in "also remember this."

The mailbox is not memory. It is a work log. Do not copy it into Hermes memory by default.

17. Secrets

  • Generate API_SERVER_KEY on the host. Store under that hostname in the vault.
  • SSH private keys stay on the Grok Bot computer, mode 600.
  • Messaging tokens are per-profile. Do not reuse across homogenized hosts.
  • Mailbox token: narrow, repo-scoped, planted by the human in the agent's vault. Never an admin's credential.
  • Webhook URL and sender key: copied by the human from the routine panel into the pinger's vault. The bot never sees the key.
  • Nothing secret goes in chat, git, or the mailbox.
  • If a key has been in a screenshot or a group chat, rotate it.

18. First-host build

  1. Mac: hermes -z "PONG" works locally.
  2. Mac on tailnet. sshd running.
  3. Grok Bot computer on tailnet. Sees the Mac.
  4. Key auth + BatchMode SSH prints PONG.
  5. Clone bridge. Add stdio MCP.
  6. Create identity bot. Confine rules.
  7. Bot: health, then PONG.
  8. Only then enable :8642, bind to Tailscale IP, vault the key, curl /health.
  9. Add hermes-<hostname> connector. Repeat PONG over HTTP.
  10. Second Mac: repeat. Then Fleet Ops + COS.
  11. Create the mailbox file in a git repo. CoS posts mailbox open.
  12. Human plants a narrow mailbox token in Regis's vault. Regis posts write test. Read back.
  13. Set CoS and Regis polls (§12).
  14. Create the CoS webhook routine. Human plants URL and header in the pinger's vault.
  15. Pinger sends one test POST after a mailbox post. Confirm CoS wakes and replies in the mailbox.
  16. Only after pings are proven, drop the poll to hourly.

19. Cadence

  • Morning: Fleet Ops health. Quiet unless broken.
  • Day: COS for work. Identity bot only when you need that host's hands. Asks over :8642. SSH for surgery.
  • Day: mailbox polled every 15 minutes, 8am to 7pm local (hourly once webhooks are proven). Webhook pings wake CoS in between.
  • Weekly: one update pass, secrets discovery, memory health (fail loud).

20. Lessons learned

Lesson What broke Rule now
Evidence for "done" A close-out letter claimed work never done; it was approved as written Any "done" claim points to a link, screenshot, or log line before CoS approves
Stale "current text" A draft quoted old text from a non-live source Read the live page on the same surface before drafting a replacement
Vault 401 The agent's vault credential failed; every write needed a hand-off token Human plants one narrow repo token in the agent's vault
Approval claimed early Regis said "Gatekeeper approved" before CoS had a record Approvals count only when logged by CoS or relayed by the Gatekeeper
Someone already fixed it The target text was already live when the edit was about to run Re-check the live page right before the edit; stand down if done

What worked: a self-correction in the mailbox plus one go/no-go question kept it to one touch for the Gatekeeper.

21. Worked example

VeriPath running a client punchlist for Phoenix (a salon). Illustration only.

  1. Gatekeeper GO. CoS posted a Job Card: work the punchlist one page per session, Google Business Profile (GBP) description first.
  2. Regis posted current vs proposed text. CoS ACKed and corrected the "current" (it was stale).
  3. Before pasting, the human read the live GBP. The description was already live at 691 of 750 characters, written by someone else.
  4. Regis put the paste on hold and posted the live text. CoS fact-checked it: KEEP AS-IS. No paste, no Save.
  5. Next: Regis logs the page done and names the next page.

22. Troubleshooting

Symptom Likely cause Fix
SSH timeout Off tailnet / ACL tailscale status both sides
Ping works, SSH hangs Userspace Tailscale ProxyCommand ts nc
hermes: command not found Non-interactive PATH HERMES_RELAY_HERMES_BIN=$HOME/.local/bin/hermes
:8642 refused Still bound to 127.0.0.1 Set API_SERVER_HOST to Tailscale IP, restart gateway
HTTP 401 Wrong key Use that host's API_SERVER_KEY
Other bots calling Scooby Account-scoped MCP Tighten prompt; remove connector from other bots
Memory filling with Grok summaries Write path left open Strip write instructions
Mailbox write 401 Wrong or unplanted token Human replants the token in the vault. Never paste it.
Mailbox write conflict Stale sha Re-read, re-append, write again
No wake after a ping Bad URL or header in the vault; non-2xx response Check vault entries, check the response code, fall back to poll
Burst of pings Misfire CoS stays quiet; mentions it once
Agent says "I don't see it" Relay text never posted to the mailbox Post it in the mailbox
Draft replaces text that is not live Stale source Re-read the live page; redraft
tailscale ping to Regis fails By design; Regis is outbound-only Not a fault. Use the mailbox.

Debug order: local -z → tailscale ping → BatchMode SSH PONG → MCP health → Grok prompts → mailbox read → mailbox write + read back → ping response code → CoS routine log → poll fallback.

23. Security checklist

  • Tailscale only. No Funnel. No 0.0.0.0 on :8642.
  • ACL: controller → 22 / 8642 / 9119 on Hermes nodes.
  • Unique API_SERVER_KEY per host, vaulted.
  • SSH key-only, BatchMode, IdentitiesOnly.
  • Identity bots confined.
  • Grok does not write Hermes memory plugins.
  • Rebuild of the Grok Bot VM must not require rebuild of Hermes.
  • Mailbox holds no secrets. Evidence by path.
  • Mailbox token is narrow and repo-scoped, planted by the human. Never an admin's credential.
  • Outside the app, the webhook URL and header live only in the pinger's vault. The bot never sees the key.
  • Ping body is a doorbell. CoS acts only on mailbox content.
  • Pinger is outbound only. No new open ports on Hermes hosts.
  • Regis: outbound only, /health only, no sshd, no logins, no sends.
  • Human does every login and every final Save / Send / publish / spend.
  • Rotate any key seen in chat or a screenshot.

24. Sources

Hostnames, URLs, keys, tokens, and ACL tag names in examples are placeholders.

Reference in New Issue View Git Blame Copy Permalink