Claw Starter

Minimal self-modifying Claw starter. Telegram, email, browser automation, and a cron scheduler / gateway. Works with Claude Code, Codex, Gemini CLI, and anything that reads CLAUDE.md or AGENTS.md.

You bring your own agent runtime (Claude Code, Codex, etc.) — we don't own or wrap it. This is just minimal scaffolding: identity files, a few tool CLIs, and a scheduler that invokes the agent on a cron or incoming message. The assistant edits its own config, memory, and personality as it works.

Quick Start

# Clone this repo anywhere
git clone https://github.com/kzahel/claw-starter.git
cd claw-starter && npm install

# Create an instance (private data, lives outside this repo)
mkdir -p ~/.assistant-data/assistants/dave
echo 'name: Dave' > ~/.assistant-data/assistants/dave/config.yaml
npx tsx lib/build.ts --instance ~/.assistant-data/assistants/dave
cd ~/.assistant-data/assistants/dave && claude

Layout

Engine (this repo) — shared capabilities:

skills/          # SKILL.md definitions (browser, gmail, telegram, reddit, git-sync)
lib/             # Tool CLIs + build system
templates/       # CLAUDE.md template

Instance (private, per-assistant) — identity + state:

config.yaml      # name, skill config, schedules (required)
soul.md          # personality
user.md          # who the user is
projects.md      # project context
.env             # secrets
memory/          # learnings, activity logs (runtime)
state/           # skill checkpoints (runtime)
CLAUDE.md        # compiled output (generated)
AGENTS.md        # symlink to CLAUDE.md

build.ts merges engine skills/template with instance files into a compiled CLAUDE.md. The agent knows when to rebuild itself.

Skills

Skills are SKILL.md files — markdown instructions, no runtime coupling. Works across Claude Code, Codex, Copilot, Cursor, etc. Here's the starter pack:

Skill	Description
browser	Headless Chromium automation
gmail	Email via IMAP/SMTP
telegram	Telegram Bot API messaging
reddit	Subreddit scraping (uses browser)
git-sync	Dirty tree / unpushed commit detection

Invocation

# Interactive (default)
claude

# Scheduled task
ASSISTANT_TRIGGER=cron:morning-digest claude -p "Run morning-digest"

# Messaging channel
ASSISTANT_TRIGGER=channel:telegram claude -p "User says: ..."

Key Decisions

• No agent loop. The provider (Claude Code, Codex) handles API calls, tool execution, reasoning. We just provide tools and context.
• Skills are prompts, not plugins. No loader, no lifecycle hooks. A SKILL.md works in any agent that reads markdown.
• Fully self-modifying. The assistant can edit its own files. Git history is the audit trail.
• Two repos, clear boundary. Engine = capabilities (public). Instance = identity (private).

Docs

• Getting Started — Set up Telegram + Gmail in 10 minutes
• docs/design/ — Format spec, comparisons with other projects, schemas, architecture decisions