Compabob

English | Español

A ready-made AI work assistant built on Claude Code. Download it, answer a few questions, and you have an assistant that knows your role, remembers what you tell it, keeps your notes in a knowledge base it can search, and never sends anything without your sign-off. No programming needed: if you can install an app and copy a link, you can set this up.

What you need

• A paid Claude plan: Pro or Max. The free claude.ai plan does not include Claude Code. See claude.com/pricing. Compabob itself is free and has no account of its own.
• A Mac or a Linux computer. Windows works through WSL, see Windows.
• About 15 minutes the first time, most of it waiting for downloads.

Choose how you want to use it

Claude Code runs in three places. Compabob works the same in all of them. Pick one:

Path	Best for	Terminal needed?
A. Claude desktop app	Most people. You may already have the app.	No
B. VS Code	People who like several conversations side by side, or who want to see and edit the notes in one window.	No
C. Terminal	People comfortable typing commands.	Yes

Every path has the same four stages: install Claude Code, get the Compabob folder, run setup, start talking.

Path A: the Claude desktop app

1. Install the app. Download Claude for Mac, install it, open it, and sign in with your Pro or Max account. Claude Code lives in the Code tab at the top of the app's sidebar. (On Windows, see Windows first.)

2. Get the Compabob folder. In the Code tab, look at the row of controls under the prompt box. Click the first one (the environment) and pick Local. Click the one next to it (the project folder), choose Open folder…, and pick your home folder (the one with your name, where Documents and Downloads live). First give your Mac the free developer tools Compabob needs (they include git and Python). Type:

   Please install Apple's developer tools with xcode-select --install.

   Claude asks permission to run the command: choose Yes (some windows call it Allow). A window from Apple opens, possibly behind the Claude window: click Install, accept the license, and wait until it says the software was installed (5 to 30 minutes). If Claude says the tools are already installed, that is fine. Then type:

   Download https://github.com/chacosoldier/compabob into a new folder called compabob here, using git clone.

   Choose Yes again.

   Prefer a download? On the GitHub page click the green Code button, then Download ZIP. Open the ZIP from Downloads, rename the folder to compabob, and drag it into your home folder (in Finder: Go > Home). It works the same, but a ZIP copy cannot pull updates later.

3. Start a new session in the compabob folder. The folder is fixed once a session has started, so start a new session from the sidebar. Before typing anything, use the project folder control again: Open folder…, then the new compabob folder.

4. Run setup. Type /setup and press Enter. Your assistant asks a few questions right in the chat (its name, your name, your role, your language, the kind of work you do). Answer in plain words. One permission box appears for the setup script: choose Yes. When it finishes, your assistant introduces itself by its new name and suggests a first task. From the next session on, it loads its memory automatically.

Coming back tomorrow: open the Claude app, go to the Code tab, and pick a session in the compabob folder from the sidebar, or start a new one there.

Path B: VS Code

1. Install VS Code and the extension. Download VS Code and open it. Press Cmd+Shift+X (Mac) or Ctrl+Shift+X (Linux), search for Claude Code (publisher: Anthropic), and click Install. The extension includes Claude Code itself.
2. Get the Compabob folder. Press Cmd+Shift+P (Mac) or Ctrl+Shift+P (Linux) to open the Command Palette, type Git: Clone, press Enter, and paste https://github.com/chacosoldier/compabob.git. Choose your home folder as the location, then click Open when VS Code offers to open the cloned folder. If VS Code says git is missing, accept its offer to install it (on a Mac: the "command line developer tools"), then repeat this step.
3. Open Claude and run setup. Click ✻ Claude Code in the bottom-right corner of the window (or the Spark icon in the left sidebar). The first time, click Sign in and finish in your browser. Then type /setup and press Enter, answer the questions, and choose Yes on the one permission box for the setup script.
4. Start talking. Your assistant introduces itself by its new name and suggests a first task. From the next conversation on, it loads its memory automatically.

Coming back tomorrow: open VS Code. It reopens the compabob folder; if not, use File > Open Recent. Click ✻ Claude Code to continue.

Tip: in VS Code you can also open vault/ in the file list on the left and read or edit your notes next to the conversation.

Path C: the terminal

Every step: paste the command, press Enter, wait for it to finish.

1. Open the Terminal. On a Mac press Cmd + Space, type Terminal, press Enter.

2. Install Homebrew (it installs the other tools). It needs an administrator account on the Mac; if it says you are not one, ask whoever set up the computer.

   /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
   

   It asks for your Mac password (nothing appears as you type, that is normal). When it finishes, it prints "Next steps" with two lines to run so your terminal can find Homebrew. Copy those exact lines. On Macs from 2020 onward they look like this:

   echo 'eval "$(/opt/homebrew/bin/brew shellenv)"' >> ~/.zprofile
   eval "$(/opt/homebrew/bin/brew shellenv)"
   

3. Install the tools Compabob uses.

   brew install git python uv node
   

4. Install Claude Code.

   curl -fsSL https://claude.ai/install.sh | bash
   

   The official install guide covers anything unusual.

5. Download Compabob into your home folder.

   cd ~
   git clone https://github.com/chacosoldier/compabob.git
   cd ~/compabob
   

6. Run setup. It asks a few short questions. Press Enter to accept the suggestion in brackets.

   ./setup.sh
   

7. Start your assistant.

   claude
   

   The first time, it opens your browser to sign in. If the browser does not open, press c to copy the link and paste it into your browser. Then type hi and press Enter: your assistant greets you by name and suggests a first task.

   If it says claude: command not found, close the Terminal window, open a new one, run cd ~/compabob, and try again.

Coming back tomorrow: open the Terminal and run cd ~/compabob, then claude (a new conversation, with memory loaded) or claude --continue (pick up the last conversation).

Linux

Skip Homebrew. Install the tools with your package manager, then follow Path C from step 4 (or Path B, which works the same on Linux):

• Ubuntu, Debian: sudo apt install git python3 curl nodejs npm, plus uv for the optional integrations.
• Fedora: sudo dnf install git python3 nodejs. Arch: sudo pacman -S git python nodejs npm.
• Check python3 --version shows 3.9 or newer before running setup.
• If ./setup.sh says "Permission denied", run chmod +x setup.sh first.

Windows

Install WSL (in PowerShell: wsl --install, then restart), open the Ubuntu app it adds, and follow the Linux steps. Claude Code also installs natively on Windows, and the Windows desktop app can run sessions inside WSL, but Compabob's safety checks are bash and Python scripts that are only tested in a WSL terminal.

What you will see the first time

• A question about trusting this folder. Choose yes. It is your own copy of Compabob, and the trust is what lets its settings and safety checks load.

• Claude waits for you to type first. Say hi. If setup has not run yet, your assistant says so and offers /setup. After setup, it greets you by name and suggests a first task.

• Permission boxes. Before Claude runs a command or changes a file, it can ask you first. Compabob is set up so that saving notes into your knowledge base (vault/), its memory (memory/) and your settings file never asks, while running a command always asks. The buttons:

  • Yes (sometimes Allow): allow this one action.
  • Yes, and don't ask again: for file edits, stop asking until the session ends. Commands still ask every time in Compabob; that is on purpose.
  • Yes, and switch to auto mode: let Claude decide which routine actions need no box. Commands still ask.
  • No: Claude stops and asks what to do instead.

  If you are unsure what a command does, choose No and ask "what would that command do?"

• A "Style note" line after some replies. Compabob checks that answers stay readable. To turn the note off, say "turn off the style notes".

• Your files. Everything lives in the compabob folder in your home folder: vault/ holds your notes, memory/ what the assistant remembers about you. Both are plain text files you can open and edit. To browse the notes as a linked map, open vault/ in Obsidian (Open folder as vault).

• Leaving. Close the window, or type /exit in the terminal. Nothing is lost: notes and memory are saved as you go.

Something looks wrong? Ask your assistant: "run the health check" (it runs bash scripts/init.sh and explains the result). A "warning" there is an optional item, not a breakage.

Your first hour

Get one loop working before you change anything:

1. Paste notes from a real meeting and say "save this to the vault as a meeting note".
2. In a new conversation, ask "what do we know about [that person or topic]?"
3. Ask "draft a follow-up email using that context". You get a draft, never a sent message.

Then try /morning-brief or /meeting-prep. The onboarding guide has the first-week checklist and what to personalize next.

Updating

Ask your assistant "update the kit" (in a terminal: ./update.sh in the compabob folder). Your data (vault/, memory/, config/user.config.yaml, .mcp.json, .env) is never touched by an update. If you changed one of the kit's own files and the update changes the same lines, the update stops, names the files that need a decision, and keeps a copy of your changes. Your assistant walks you through it. See the customization guide.

What is inside

CONSTITUTION.md         The rules every session loads. How your assistant behaves.
.claude/agents/         Specialized agents (analyst, comms, CRM, strategy, ...).
.claude/skills/         Slash-command workflows (/setup, /morning-brief, /meeting-prep, ...).
.claude/output-styles/  "Direct", the answer-first response style.
.claude/settings.json   Which actions ask for permission, and the safety hooks.
hooks/                  Safety guards and the session-start briefing.
memory/                 What the assistant remembers (yours, created by setup, git-ignored).
vault/                  Your knowledge base, an Obsidian vault (yours, git-ignored).
config/                 Persona presets (kit) and your user.config.yaml (yours, git-ignored).
modules/                Opt-in extensions (see below).
scripts/                The health check, the integrations installer, shared helpers.
setup.sh, update.sh     One-time setup, and pulling the latest kit.
docs/                   Architecture, onboarding, and customization guides.
tests/                  The checks CI runs on every change.
*.example/              Seeds that setup copies into vault/ and memory/.

Core agents

Agent	What it does
daily-copilot	Your everyday partner: prioritization, sparring, the daily brief
second-brain	Knowledge base operations: notes, meeting records, research synthesis
analyst	Metrics, KPIs, dashboards, data-backed reporting
crm-relationships	Contact, pipeline, and relationship tracking (works with or without a CRM)
comms-meetings	Email triage, meeting prep, briefings, follow-up tracking
strategy-advisor	Assumption testing, pre-mortems, second-order effects (advice only)
principal-engineer	Architecture and code review for technical work
first-principles	Rigorous reasoning under uncertainty
skeptic	Checks a claim or finding against the files before anyone acts on it (never edits)

Requests route automatically. You do not pick an agent.

Skills

Skills are slash-command workflows. Type the command; the assistant runs the procedure.

Skill	What it does
/setup	One-time setup from inside the chat: a few questions, no terminal
/morning-brief	Start-of-day briefing: priorities, what is due, what is owed
/meeting-prep	Assemble attendees, history, and open items before a meeting
/post-call	Capture decisions and commitments right after a call
/handover	Write a handover note so the next session has full context
/log-decision	Record a decision, its reasoning, and a date to review it
/council	Put a hard-to-reverse decision to an advisor panel with a mandatory dissenter, then answer each objection
/design-precheck	Pre-mortem a big plan with three advisors before any work starts
/tasks	One aggregated view of open tasks from across the vault
/reflect	End-of-session reflection; proposes memory updates
/index-memory	Build the memory-search index for relevance-ranked retrieval
/add-agent	Scaffold a new specialized agent interactively
/system-audit	Health-check the assistant's own setup
/visual-explainer	Turn a system, plan, or dataset into a self-contained HTML page
/document-export	Create a PDF, Excel, or Word file from your content
/weekly-review	End-of-week review: done, slipped, decisions, commitments, next week's top 3
/chart-tufte	Self-grade any chart against Tufte's Visual Display of Quantitative Information before showing it
/mcp-debug	Health-check, trace, or audit your MCP servers when tools fail silently
/memory-debt	Resolve memory updates that earlier sessions proposed but never wrote
/merge-contacts	Fold your Google + LinkedIn + vault contacts into one local CRM, deduped
/build-list	Build a ranked outbound list: discover, dedup against your CRM, enrich, score

Modules (opt-in)

The core runs with zero external services. Modules add capability when you want it:

Module	Status	What it adds
proactive	Available	Scheduled automation: a morning brief and weekly review on a timer
telegram	Available	A Telegram bot: inbound messages drafted for your approval, never auto-sent
integrations	Available	MCP tools: browser automation, web search, utilities. Gmail and Calendar via Claude's own connectors
linkedin-outreach	Available	Drafts one LinkedIn connection-invitation card a day from a queue, for your review and manual send
crm-merge	Available	Folds Google + LinkedIn + vault contacts into one local CRM (SQLite + offline HTML browser), deduped across sources. No creds
lead-pipeline	Available	Builds a ranked outbound list: discover, clean, dedup against your CRM, enrich, score. One CSV per stage
memory-search	Available	Real index over memory/ and vault/ so retrieval ranks by relevance (FTS5 by default; semantic via Ollama if installed)
transcribe	Available, highly encouraged	One-hotkey local call recording + transcription (mlx-whisper). Captures both sides, copies to clipboard, files a transcript into vault/raw/meetings/. No cloud, no key, audio stays on your machine
dictation	Available	Cleanup endpoint for your dictation app: fixes punctuation and fillers, and a glossary that learns the words it keeps getting wrong

See modules/README.md to enable one, and for what is deliberately not built yet (an agent gallery, team mode, WhatsApp).

Choosing the Claude model

Compabob follows whatever model Claude Code is set to. Every agent uses model: inherit, so when you switch models, the agents switch with you. Claude Code's default today is Opus 5.5 on every paid plan.

• Per session: type /model and pick from the list (the desktop app and VS Code have a model picker under the prompt box), or start the terminal with claude --model sonnet.
• Per project: add a "model" field to .claude/settings.json. Aliases such as "opus", "sonnet", "haiku" work, or a full model ID such as "claude-sonnet-5-5".
• Globally: set the ANTHROPIC_MODEL environment variable.

On a Pro plan, the larger model uses up your limits faster, and agents that inherit it do too. If you hit limits, switch to sonnet with /model, or pin the agents: change model: inherit to model: sonnet in the files in .claude/agents/. The scheduled modules (proactive, telegram, linkedin-outreach) already run on Sonnet. If you use the assistant every day, or run several conversations at once, Max removes most of that ceiling.

Current model IDs are listed in the model overview.

Why

A fresh Claude Code session is a strong generalist with no memory of you. Every time, you re-explain who you are, how you want answers, and what you are working on. It has no sense of your priorities, no guardrails on what gets sent, and no place to keep what it learns.

Compabob is the missing structure. The core loop, on day one:

You ask it to take notes from today's planning meeting. It files them into your knowledge base. Next week, before the follow-up, you type /meeting-prep and it hands back the attendees, what was decided, and the open items, without you searching for anything.

That second-brain loop is the value you get immediately. Compabob is scaffolding, not a finished assistant: everything else compounds as you make it yours.

What this is (and is not)

• Is: an opinionated, single-user template. Architecture, conventions, and a set of agents, hooks and skills that took real iteration to get right, packaged so you do not start from zero.
• Is not: a product, a SaaS, or a no-config magic box. There is no signup, no telemetry, no upsell. It runs entirely on your machine under your own Claude subscription.

How this differs from the other Claude Code things you have seen

• vs. raw Claude Code. Claude Code is the engine. Compabob is the opinionated setup on top: a constitution, an agent fleet, safety hooks, a memory system, and a knowledge base, all wired together. Without it, every new session starts from zero.
• vs. "awesome" lists (curated link directories). Those tell you what exists. Compabob is a working kit you download and run.
• vs. multi-agent dev-team frameworks. Those assume you are building software with AI. Compabob is built for a single knowledge worker doing their actual job: meetings, notes, research, comms, decisions.
• vs. building your own from scratch. That is the alternative most people pick, and most never finish. The structure here took real iteration; you skip that part.

Maintenance posture

Actively maintained. The kit is updated as its maintainers learn what works and as Claude Code evolves. You still own your copy outright: change anything, diverge as far as you like. Issues and PRs are welcome; reviews are best-effort, since the kit is maintained alongside other work.

Feedback

Found a bug or a rough edge? Open an issue. For questions, setup help, or general discussion, use GitHub Discussions. Improvements are welcome as pull requests, see CONTRIBUTING.md.

Lineage

This kit is distilled from a personal Claude Code setup the author refined over a year of daily use: the constitution, the agent fleet, the safety hooks, the memory system. The personal-life domains were stripped out and the patterns generalized into a clean starting point. Credit to the Claude Code team for the engine this builds on.

Author

Built by Philipp Wenger Lebron. Issues and pull requests are welcome.

License

MIT. See LICENSE. Use it, change it, ship it.