Clawford

Frozen historical guide. This is the OpenClaw-era field manual, kept here as a reference and a record of the migration. The live guide is in guide-v3/. New readers should start there.

Anything in this document that mentions the OpenClaw runtime, the Docker gateway, the oc / oci CLI wrappers, exec-approvals, or ~/.openclaw/ paths is describing how Clawford worked before Phase 7 of the liberation (April 2026). Useful for archaeology, not for setting up a new fleet.

What Is Clawford?¶

Last updated: 2026-04-13 · Reading time: ~8 min · Difficulty: easy

TL;DR

Clawford is a personal fleet of LLM agents running on a cheap VPS. Each agent is a teleported Busytown character — yes, the Richard Scarry ones — and each has a narrow job.
This guide is the scar-tissue version — what I'd tell a friend over beers, not a pitch deck. Most of it is written because something bit me first.
It's aimed at people who want multiple agents sharing infrastructure, not one standalone bot.
If you want the myth version before the manual version, read docs/ballad-of-mr-fixit.md first.
Start with Ch 02 — Before you start when you're ready to move.

A fox named Mr Fixit¶

Every morning at 5am, a fox named Mr Fixit checks that the lights are still on in Clawford and sends me a one-line status to Telegram. If the news agent has stopped fetching headlines, Mr Fixit tells me. If a Dropbox sync has conflicted, Mr Fixit tells me. If everything is fine — which, astonishingly, it sometimes is — he says nothing at all, which I have come to recognize as its own kind of love language.

Clawford is what I call the thing that runs him, his six coworkers, and the small file-based shared memory they all write into. The animals themselves teleported in from Richard Scarry's Busytown — I borrowed them wholesale and put them to work here, for reasons that will not become clearer. It lives on a $30/month Hetzner box, stitched together with OpenClaw, Dropbox, Telegram, some Python, and a great deal of hard-won regret.

What Clawford tries to do¶

I wanted software I could trust to do small, recurring, personal-life jobs while I slept:

Curate the news I actually read (Lowly Worm 🐛📰).
Reorder the things I keep running out of (Hilda Hippo 🦛🛒).
Keep our family calendar honest (Mistress Mouse 🐭📅).
Prep me for meetings and debrief me after (Sergeant Murphy 🐷🔍).
Remember who's who in my life, and when to follow up (Huckle Cat 🐱🤝).
Watch the whole fleet so none of the above quietly die (Mr Fixit 🦊🔧).

Each agent has a narrow remit on purpose. They share a common brain — a directory of markdown files with schemas — so one agent can hand context to another without me playing messenger.

What Clawford does not try to do¶

It is not a general-purpose assistant. It does not replace humans, make autonomous financial decisions, post to family group chats on my behalf, or hold opinions I didn't authorize it to hold. When an agent is unsure, it asks me. When it's sure and wrong, it eventually finds out, and I have words with it in the form of a new entry in a MEMORY.md file.

The shape of the thing¶

          ┌─────────────────────────────────────────┐
          │          Telegram (you ↔ fleet)         │
          └───────────────────┬─────────────────────┘
                              │
                ┌─────────────┴────────────┐
                │     OpenClaw gateway     │   ← LLM reasoning,
                │     (Docker, on VPS)     │     exec gating, routing
                └─────────────┬────────────┘
                              │
    ┌────────┬───────┬────────┼────────┬────────┬────────┐
    │Fix-It  │Lowly  │ Hilda  │ Mouse  │ Murphy │ Huckle │  ← agents:
    │ 🦊🔧   │ 🐛📰  │  🦛🛒  │  🐭📅  │  🐷🔍  │  🐱🤝  │    SOUL · IDENTITY
    │        │ Worm  │  Hippo │        │        │  Cat   │    · TOOLS · crons
    └────────┴───┬───┴────────┴────────┴────────┴────────┘
                 │
          ┌──────┴──────────┐
          │  Shared brain   │   ← Dropbox-synced markdown
          │ (facts, people, │     (decays, schemas, IDs)
          │  tasks, notes)  │
          └─────────────────┘

Two kinds of work run under that gateway. LLM reasoning — ranking, composing, judgment calls — runs inside OpenClaw crons. Deterministic work — scrapers, browser automation, fleet-health probes — runs as plain Python or shell, either in-container or as host crons on the VPS. Ch 06 draws that line more carefully.

VPS or Mac mini?¶

I went with a VPS for four reasons, none of them religious:

Always-on without kicking a family laptop. A Mac mini under the TV works until someone unplugs it to dust.
Blast radius. If an agent goes feral, I'd rather that happened on rented hardware than on the box with my tax returns on it.
Clean rebuild. The whole fleet is Terraform + Docker, so "try again from scratch" is a 40-minute thing, not a weekend.
Not my residential IP. For most services this is a feature; for Amazon and Costco it is the exact opposite, which is how I ended up with strong opinions about residential proxies (Ch 03).

A Mac mini is probably fine if your fleet is smaller, you already own the hardware, and you don't mind the first three points. I don't claim one path is right.

Who this guide is for¶

You're comfortable driving SSH, git, shell, and Python with Claude Code or Codex as a copilot. You don't need to be an expert in any of it — you need to be able to wield the tools well enough to unstick yourself when the LLM runs out of ideas.
You want to run multiple agents sharing infrastructure. The whole guide is written around that case.
You're OK with "works on my box" as a starting point and willing to carry some operational weight yourself.

Who it isn't for¶

You just want one bot. If all you need is a single agent to do one thing, Busytown is massive overkill. Clone a smaller scaffold.
You want a turnkey product. This is a field guide, not a product. There are places in it where the answer is "I gave up and wrote a memory file."
You need a security-reviewed system. I'm not a security expert. I'll tell you what I do and what I know I don't do (Ch 08), but audited it is not.

If you read nothing else — quickstart¶

Sign up for a Hetzner Cloud account and a Dropbox account you're willing to dedicate to this.
Create a Telegram account and start a chat with @BotFather — you'll make bots there.
Clone the repo and read Ch 02 — Before you start end-to-end before you provision anything.
Follow Ch 03 — VPS setup to stand up a locked-down box.
Walk Ch 05 — Infra setup to wire up the shared brain, Dropbox daemon, and Telegram gateway.
Deploy Mr Fixit first, following Ch 07-1. He's the training-wheels agent and the canary for everyone after him.
Wait for your first 5am status message. When it arrives, pour yourself something, and then pick your second agent from Ch 07.

Caveat emptor¶

I'm writing this as I learn. Half of it is scar tissue from mistakes, and some of it will be wrong by the time you read it — services change their HTML, APIs deprecate, cookies mutate, and every so often I do something clever with a shell script at midnight that I regret by breakfast. Where I can, I've tried to tell you why a thing is the way it is, so that when it stops being that way you'll have enough context to adapt. Where I can't, I've just told you what bit me and left it at that.

If you find something here that's wrong, assume I'd like to know.

Full table of contents¶

#	Chapter	What you'll do	Reading	Difficulty
01	What is Clawford?	(you are here) — orient, decide if this is for you	~8 min	easy
02	Before you start	Prerequisites, costs, the choices that save time later	~15 min	easy
03	VPS setup	Provision and lock down a Hetzner box, Tailscale, residential proxy	~25 min	moderate
04	Dev setup	Claude Code, test harness, the limits of trusting the robot	~20 min	moderate
05	Infra setup	Shared brain, Dropbox, Telegram gateway, unified deploy tool	~30 min	moderate
06	Intro to agents	Anatomy of an agent: SOUL, IDENTITY, TOOLS, crons, script contract	~20 min	moderate
07-0	Your first agent	Deployment order and the general first-deploy arc	~10 min	moderate
07-1	Mr Fixit 🦊🔧	Infra, monitoring, brain validation, the first-deploy minefield	~30 min	hard
07-2a	Lowly Worm — newsfeed 🐛📰	News curation and the LLM-judge preference learning that makes it actually useful	~20 min	moderate
07-2b	Lowly Worm — social channel extension 🐛📰 (optional)	LinkedIn notifications + DM consolidation, the dark-art scraping story	~25 min	hard
07-3	Mistress Mouse 🐭📅	Google Calendar, reminders, family logistics	~20 min	moderate
07-4	Sergeant Murphy 🐷🔍	Meeting prep, Workflowy sync, coaching	~25 min	moderate
07-5	Huckle Cat 🐱🤝	Relationship memory, notes triage, social awareness	~25 min	moderate
07-6	Hilda Hippo 🦛🛒	Amazon and Costco as representative purchasing flows	~35 min	hard
07-7	Auth architectures	What works, what doesn't, what's still brittle	~20 min	moderate
08	Security and hardening	Attack surface, defense in depth, what I don't do	~20 min	moderate
09	Scripts and configs reference	Categorical and alphabetical index of every moving part	bookmark	reference
10	CLI reference	Task-indexed `oc` / `oci` commands that actually work	bookmark	reference
11	Glossary	Every bolded term, in one place	bookmark	reference

Reading times and difficulty ratings are placeholders until each chapter's header stamp lands — I'll reconcile them once the rest of the draft is in.