Hermes Autopilot Kit

Free, open-source playbook and one-command installer for building a self-improving Hermes agent from scratch.

Use this when you want Hermes running with a private brain, receipts, starter jobs, and a self-improvement loop without hand-building the stack.

One-Command Start

Run this on a VPS or local machine with shell access:

curl -fsSL https://raw.githubusercontent.com/amarrtech/hermes-autopilot-kit/main/install.sh | bash

The installer creates:

• persistent Hermes data directory
• Docker Compose service using the official nousresearch/hermes-agent:latest image
• markdown-first BrainCache repo
• GBrain indexing when ENABLE_GBRAIN=true
• receipt writer
• gateway health check
• brain sync job
• weekly self-improvement scorecard
• local cron entries for starter checks

Then it prints the exact command to finish Hermes provider and gateway setup:

docker run -it --rm -v "$HOME/.hermes-autopilot/data:/opt/data" nousresearch/hermes-agent setup

On a fresh install, the script launches this setup wizard automatically before starting the gateway. You only rerun it later if you want to change provider, model, or chat gateway settings.

Promise

Hermes Autopilot is a private ops agent that:

• watches important workflows
• sends useful Discord or Slack reports
• saves durable memory to a markdown brain
• writes receipts for every run
• notices repeated failures
• improves its own skills and playbooks over time

The user should only provide credentials. The AI agent should do the rest.

User Inputs

Minimum:

• VPS or local machine shell access
• Discord bot token or Slack app token
• model provider key or login method
• GitHub repo for backups

Optional:

• domain name
• Cloudflare tunnel
• X API / xurl credentials
• Google Workspace credentials
• Stripe key for spend/watch workflows

Never ask the user to edit Docker files, cron files, or YAML by hand.

What Is Still Manual

The installer cannot create third-party accounts for you.

Users still need to create or provide:

• Discord or Slack app/token
• model provider auth
• GitHub backup repo/token
• optional Cloudflare, X, Google, or Stripe credentials

Everything after credential collection should be handled by the installer or by an AI agent using AGENT_BUILD_PROMPT.md.

GBrain Behavior

GBrain is default-on.

The installer tries to:

1. install Bun
2. install gbrain
3. run gbrain init --pglite
4. sync BrainCache markdown into GBrain
5. add a 5-minute brain sync cron
6. run gbrain doctor

If any GBrain step fails, Hermes Autopilot still works in markdown-only mode. The markdown brain is the source of truth; GBrain is the searchable index that can be repaired or rebuilt later.

Disable GBrain during install by answering:

Install GBrain indexing? false

Target Architecture

User
  |
  v
Discord / Slack
  |
  v
Hermes Agent
  |-- skills/
  |-- cron jobs
  |-- logs
  |-- receipts
  |
  v
GBrain / BrainCache
  |-- markdown notes
  |-- daily memos
  |-- scorecards
  |-- contradiction reports
  |
  v
GitHub Backup Repo

What The AI Agent Must Build

1. Runtime

• Install Docker or use existing Docker.
• Create persistent data directory.
• Run Hermes in a container.
• Start Hermes gateway for Discord or Slack.
• Add a watchdog so gateway restarts if it dies.
• Run hermes doctor and fix basic issues.

2. Brain

• Create a separate markdown repo, for example BrainCache.
• Install GBrain or compatible markdown-first brain.
• Store all durable knowledge as markdown first.
• Treat the database/vector index as rebuildable.
• Sync markdown into GBrain on a schedule.

3. Skills

Install these starter skills:

• takeaway — extract durable lessons from links and videos
• topic-synthesis — summarize everything the brain knows about a topic
• self-improvement-engine — inspect failures and improve playbooks
• brain-ops — search, write, and repair brain notes
• backup-ops — verify GitHub backups
• gateway-health — check Discord/Slack gateway health
• competitor-watch — monitor products or companies
• content-drafts — draft X/LinkedIn posts, never auto-post without approval

4. Default Jobs

Create these recurring jobs:

Job	Frequency	Agent?	Output
Gateway health check	every 15 min	no	alert only if down
Brain sync	every 5 min	no	GBrain sync log
Daily signal memo	daily	yes	brain page + chat summary
Weekly self-improvement	weekly	yes	scorecard
Weekly contradiction scan	weekly	no or yes	contradiction report
Backup check	weekly	no	GitHub backup receipt
Competitor watch	weekly	yes	product notes
Content draft review	manual	yes	draft files only

Use no_agent for deterministic checks. Wake the LLM only when a result needs judgment.

5. Receipts

Every job must write a receipt:

---
type: receipt
job_id: <stable-id>
status: ok | failed | partial
created_utc: <iso timestamp>
---

# <Job Name> Receipt

- Trigger:
- Inputs:
- Files changed:
- Checks passed:
- Checks failed:
- User attention needed:
- Next run:

Receipts go in:

brain/receipts/YYYY-MM-DD/

6. Self-Improvement Loop

The loop runs weekly:

1. Read failed receipts and logs.
2. Count repeated mistakes.
3. Detect noisy jobs and useless alerts.
4. Find stale skills or missing verification.
5. Propose small fixes.
6. Apply safe fixes automatically.
7. Ask before risky fixes.
8. Write a scorecard.

Scorecard fields:

repeat_mistakes: 0
failed_jobs: 0
noisy_alerts: 0
skills_changed: 0
brain_pages_added: 0
contradiction_candidates: 0
backup_status: ok

7. Safety Rules

• Draft before posting.
• Read-only by default.
• Mutating tools require explicit approval.
• Store secrets outside Git.
• Redact logs before backup.
• Every recurring job must have an owner, purpose, tool scope, and rollback path.

One-Click Product Shape

The current open-source version is one-command, not hosted one-click.

For a hosted version, expose this flow:

1. User clicks Deploy Autopilot.
2. User connects Discord or Slack.
3. User connects GitHub backup repo.
4. User chooses model provider.
5. System provisions Hermes, GBrain, skills, and jobs.
6. User gets a message: Autopilot is live. First health check passed.

The open-source installer is:

curl -fsSL https://raw.githubusercontent.com/amarrtech/hermes-autopilot-kit/main/install.sh | bash

The installer should ask only for credentials and confirmations.

Acceptance Checks

The build is not done until these pass:

hermes --version
hermes doctor
hermes gateway status
gbrain doctor
git -C ~/brain status
curl -f http://localhost:8765/health

Also verify:

• Discord or Slack receives a test message.
• A test receipt is written.
• A brain page can be created and searched.
• Backup repo receives a commit.
• A failed test job creates an alert.
• Self-improvement scorecard is generated.

Minimal First Launch

Ship only this first:

• Discord gateway
• BrainCache markdown repo
• GBrain sync
• gateway health check
• weekly backup check
• weekly self-improvement scorecard
• takeaway skill
• topic-synthesis skill

Everything else can be a later module.

License

MIT. Use, modify, sell hosted services, and share freely.