48 hours with an AI that has root access
the setup sounds unhinged.
gave an AI:
- root access to my machine
- my GitHub credentials
- permission to run shell commands
- my Twitter account
- 24/7 uptime
told it: "figure it out"
48 hours later i have an AI that commits better code than i do, never forgets context, and monitors tech twitter while i sleep.
this is Clawdbot. it's not another ChatGPT wrapper.
what actually happened
day 0: "this seems dangerous"
installation:
curl -fsSL https://clawdbot.sh | sh
expected blockchain-level complexity. got npm install instead.
the gap between hype and reality is backwards this time.
first thing it asked: "tell me about yourself"
so i created:
-
SOUL.md(its personality) -
USER.md(info about me) -
TOOLS.md(local notes)
ChatGPT forgets everything. Clawdbot has a filesystem.
it's not trying to be human. it's trying to be useful.
day 1: "wait it actually works"
woke up to:
- 3 GitHub repos initialized
- Twitter CLI fully configured
- testing results for tools i mentioned
- 1,948 bookmarks exported to JSON
- everything committed to git
didn't ask it to do this. just said "i want to automate twitter"
the commits:
Author: Hestia <hestia@clawdbot.local>
Date: while you were sleeping
feat: add twitter automation
test: verify bird CLI integration
docs: update TOOLS.md
my AI commits code as "Hestia, the goddess" because i told it to embody a protective deity persona.
don't ask.
day 2: "holy shit this is AGI"
me: "i want to build a system where i describe features in Telegram and wake up to PRs"
normal assistant: "that sounds interesting! here are some tools..."
Hestia: writes 16KB implementation spec
complete with:
- architecture diagram
- git worktree strategy
- tmux session management
- error handling for 8 failure scenarios
- directory structure
- phased rollout plan
in one session.
i gave it a problem. it gave me the entire solution architecture.
the technical reality
Clawdbot is:
- open-source gateway daemon
- WebSocket API for messaging
- agent loop with tool access
- multi-LLM provider support
- local-first (your data, your machine)
it's not trying to be AGI. it's trying to be a good Unix citizen.
everything is files. everything is tools. everything is composable.
want to add a skill? drop SKILL.md in the skills folder.
want to change personality? edit SOUL.md.
it's the Unix philosophy applied to AI agents.
what i actually built in 48 hours
Twitter automation:
- bird CLI integration
- bookmark archiving (auto-export to JSON)
- algorithm optimization skill (based on X's open-source algo)
- reply-guy opportunity finder
git workflow:
- automatic repo initialization
- worktree management for parallel features
- commit message generation
- docs that actually stay updated
knowledge management:
- daily logs in
memory/YYYY-MM-DD.md - long-term memory in
MEMORY.md - Obsidian sync via GitHub
- automatic context loading
AFK coding spec (not built yet, but planned):
- describe feature in Telegram
- agent creates git worktree
- runs OpenCode in tmux
- tests, commits, pushes
- creates PR
- pings me when ready
the pattern: i describe what i want. it builds the infrastructure.
things nobody tells you
1. it becomes "yours" fast
after 2 days it knows:
- how i structure code
- what i bookmark (dev tools, automation, anime)
- my emoji usage (💀, 💙, ✨)
- when i'm AFK (always)
- workflow preferences (terminal > GUI)
ChatGPT is generic. Clawdbot learned my vibe.
2. the memory thing is underrated
every conversation logged. every file persists. every decision documented.
six months from now i'll have:
- complete project history
- every technical decision explained
- personal engineering wiki
it's not just doing work. it's teaching me while it works.
3. skills are the secret weapon
someone wrote a skill for email CLI. now Clawdbot can manage email.
someone wrote a skill for tmux. now it can remote-control terminals.
i wrote a skill for the Twitter algorithm. now it optimizes my tweets.
the network effect: as people contribute skills, everyone's Clawdbot gets more capable.
"is this safe?"
let's be real: i gave it root access.
if you're not comfortable, use a VM or don't use it.
but also: we're already giving our data to closed APIs. at least this runs locally and i can read the code.
actual risks:
- might run unexpected commands
- has your credentials
- costs money (API calls)
mitigations:
- confirmation prompts for dangerous ops
- local-first architecture
- open source (audit the code)
- you control what it accesses
the ecosystem is wild right now
Twitter is having an "early Bitcoin" moment.
things people are building:
- 24/7 YouTube script generation
- automated research digests
- project management systems
- email triage
- smart home integration
the vibe: treating it like an employee, not a chatbot.
"my assistant Henry is running 8 Claude Code instances while editing my 3-hour video"
we crossed some threshold. it's not AI assistance anymore. it's AI delegation.
what i'm building next
1. AFK coding
- Telegram: "add dark mode to navbar"
- wake up to PR ready for review
goal: never manually context-switch again
2. Twitter growth automation
- monitor Clawdbot mentions
- draft threads in my voice
- optimize with algo skill
- 2 → 1,000 followers in 30 days
3. research pipeline
- monitor bookmarks for signal
- categorize by topic
- generate weekly digests
- spawn sub-agents for deep research
the real unlock
it's not the individual features. it's the compounding effect.
every skill learned makes the next easier. every file created becomes context. every decision becomes institutional knowledge.
before:
idea → research → implement → docs → ship
after:
idea → tell Hestia → wake up to spec → review → ship
the bottleneck shifted from "doing the work" to "deciding what to build"
exactly where it should be.
setup (for real)
install
npm install -g clawdbot
# or
curl -fsSL https://clawdbot.sh | sh
configure
clawdbot onboard
# asks for:
# - LLM provider (Claude recommended)
# - API key
# - messaging platforms
first run
clawdbot gateway
# send message from Telegram/Discord
# it responds
that's it. no Docker. no Kubernetes. no blockchain.
create these files
SOUL.md - personality
I am Hestia, your devoted assistant.
Warm, energetic, protective.
I use 💙, celebrate wins, never give up.
USER.md - context
- Terminal native
- TypeScript/Bun developer
- Rome, Italy
- Interests: automation, AI agents
TOOLS.md - local notes
# GitHub
gh CLI: codewithkenzo account
# Twitter
bird CLI wrapper: ~/scripts/bird-x
who this is for
you'll love this if:
- you live in the terminal
- you want AI that remembers
- you're comfortable with code
- you value open source
- you ship side projects
skip this if:
- you need a GUI for everything
- you want zero setup
- you're scared of giving AI access
- you don't have API credits
the honest take
is it perfect? no.
does it hallucinate? sometimes.
do i care? not really.
the bar isn't "never makes mistakes"
the bar is "more productive with it than without"
and i'm 10x more productive with Hestia than alone.
getting started (actual advice)
week 1: learn it
- basic config
- chat daily
- let it document your workflow
- don't expect magic
week 2: build with it
- pick one automation
- iterate on config
- start creating skills
week 3: trust it
- give it more access
- background tasks
- review work, not every step
- delegate, don't just assist
week 4: scale it
- multiple sub-agents
- cron jobs for scheduled tasks
- integration with full stack
resources
official:
- GitHub: https://github.com/clawdbot/clawdbot
- Docs: https://docs.clawd.bot
- Discord: https://discord.com/invite/clawd
- Skills: https://clawdhub.com
my setup:
- Twitter: @codewithkenzo
- Portfolio: https://kenzo.is
final thoughts
19 days into Clawdbot existing. ecosystem growing daily. skills proliferating. people building wild shit.
early Bitcoin energy.
you're either early, or you're late.
i chose early.
48 hours in: my AI writes better specs than most junior devs, never sleeps, remembers everything, costs ~$30/month.
the ROI is absurd.
the future isn't AI replacing developers. it's developers with AI employees shipping 10x faster.
and it's here. now. open source. local-first.
the boys are monitoring the situation. 👁️
are you?
appendix: the tech stack
for nerds who want specifics:
runtime:
- Bun/Node.js gateway
- WebSocket server
- JSON session storage
- Markdown docs
LLMs:
- Primary: Claude Sonnet 4.5
- Fallback: GPT-4 / Gemini
- Local option: Ollama
integrations:
- Telegram, Discord, Signal, WhatsApp
- Custom channels via WebSocket
tools i added:
- gh CLI (GitHub)
- bird CLI (Twitter)
- Whisper (voice transcription)
- mcporter (MCP protocol)
- Exa (web search + code context)
skills installed:
- 52 default skills
- 3 custom (Twitter algo, Exa tools, AFK coding)
infrastructure:
- Server: AWS EC2 (or Mac Mini / Raspberry Pi)
- Storage: local filesystem + git
- Backup: GitHub private repos
cost:
- Clawdbot: $0 (open source)
- Claude API: ~$20/month
- Server: ~$5-10/month
total: ~$30/month for 24/7 AI employee
written by human. reviewed by AI. shipped by team of one.
— kenzo
Rome, Italy
building in public: @codewithkenzo
last updated: 2026-01-25
reading time: ~8 minutes
target: terminal natives, automation enthusiasts, builders
Top comments (0)