Independent, self-contained AI agent personas built with native Claude features first. Accessible from CLI and Claude Cowork.
Hello! Personas are my take on a simple and elegant Claw-like (lol) framework using native abilities for Claude.
$ warren "how am I doing this month?"
📊 Through March 22:
You're at $4,280 of your $5,500 monthly spend target — 78%, with
8 days left. Dining is the outlier again: $620 vs $400 budget.
Savings rate: 31% Net worth: +2.1% MoM Emergency fund: 5.8 months
That Vanguard rebalance we talked about last week — rates moved
enough that it's worth doing now. Want me to draft the trades?
What's a persona?
A persona is simply a folder. Isolated from your global Claude config, backed by git, and ready to launch from any terminal. Inside:
~/.personas/warren/
├── CLAUDE.md # what the persona does, knows, and how it behaves
├── .claude/
│ ├── settings.json # sandbox, permissions, memory config
│ ├── output-styles/ # personality and tone (replaces default Claude prompt)
│ ├── hooks/
│ │ └── public-repo-guard.sh # blocks personal data leaks in public repos
│ └── skills/ # reusable workflows (self-improve ships with every persona)
├── hooks.json # session lifecycle: start, stop, compaction, git guard
├── docs/ # reference materials, plans, domain knowledge
├── tools/ # scripts, utilities, data pipelines
├── user/
│ ├── profile.md # your personal context (filled during first session)
│ └── memory/ # native auto-memory (local, git-tracked)
├── .mcp.json # MCP server connections (gitignored)
├── .claude-flags # per-persona launch flags
└── .gitignore # protects secrets; optionally ignores user/ for public sharing
Everything is created with you during setup — persona-dev interviews you, researches your domain, and scaffolds the whole thing. Keep it local or connect it to GitHub for backup. Launch
Each persona stores its launch flags in .claude-flags, configured during setup:
- Sandboxed — can only read/write its own directory
- Permission bypass — runs autonomously (only when sandboxed)
- Chrome — browser automation for personas that need it
- Remote control — accessible from Claude Desktop or other apps
Once running, personas extend themselves. They learn with native auto-memory, schedule reminders and timed checks with natural language ("remind me at 3pm to..."), and may ask to create additional files to track data or workflows. Every persona ships with a self-improvement skill — it can develop new skills, create tools, and keep itself organized with periodic audits.
Example personas
These are all examples of actual personas I created and use to help organize my busy life. In theory, a persona could be an expert assistant with pretty much whatever you want.
| Persona | Role | MCP Servers | Skills |
|---|
| warren | Personal CFO — sharp, data-driven, Buffett-esque | Monarch (finances), Google Workspace | finance, self-improve |
| julia | Personal Chef — warm, encouraging, meal planning pro | Mealie (recipes & meal plans), Google Workspace | personal-chef, self-improve |
| nara | Health & Wellness Coach — holistic, evidence-based | Consensus (research papers), HealthEx (health data), Google Workspace | wellness, self-improve |
| bob | Home Repair Expert — project scoping + how-to guides | Excalidraw (diagrams), Google Workspace | home-repair, calculators, self-improve |
You can also use ANY other Claude Code plugin or skill with a persona. Just add the marketplace as you would normally, or add them directly to the persona directory.
Quick Start
Dependencies
| Dependency | Platform | Why |
|---|
| Claude Code | All | Required — the foundation |
| bubblewrap + socat | Linux / WSL2 | Required — OS-level sandboxing (sudo apt install bubblewrap socat) |
| Git for Windows | Windows | Required — provides Git Bash runtime |
| gh | All | Optional — repo creation during setup + public repo safety checks |
Step 1: Install the plugin
Install persona-manager once — after that, every persona you create auto-installs it.
