🔬 Alpha · Go · Elastic License 2.0

KittyPaw
A local AI automation experiment

A Go project building chat, skills, scheduler, memory, and permissions in small pieces
to see how far it can go. Single binary, no external runtimes.

$ curl -fsSL https://raw.githubusercontent.com/kittypaw-app/kitty/main/install-kittypaw.sh | sh

macOS · Linux — Windows builds available on the releases page

⭐ GitHub repo Issues

kittypaw — bash

$ kittypaw setup

LLM provider: Claude (Anthropic)

Channel: Telegram

✓ Settings saved · Server up

🐾 KittyPaw started on :3000

single Go binary · sqlite (modernc) · goja JS sandbox · macOS / Linux / Windows

Components

What I'm building

Not a polished SaaS — a source-available project building local automation components piece by piece.
Each card uses ✅ / 🚧 / 🔬 to mark its current verification state.

🎯

✅ Skill packages

Install from registry / GitHub URL / local path. package.toml, source-bound config (secrets.json), SHA256 verification.

🧠

🚧 Reflection loop

Recent chats → repeated intent / topic candidates stored. First-turn suffix surface ✅ verified, "yes" reply dispatches ✅, generated code stability 🚧.

🔒

✅ Sandbox + Permission

All skills run in goja. SSRF guard, hop-by-hop header block, dangerous ops (Shell.exec / Git.push) require approval in supervised mode.

🔗

🔬 Skill chaining

Pipeline of skills. Staged execution validated for some packages. A general-purpose chain DSL is still undecided.

🌐

🚧 Web search

DuckDuckGo (no-config) + Google API + Tavily option. Web.search + Web.fetch primitives. Source quality / verification ongoing.

🏠

✅ Local LLM

OpenAI-compatible local endpoints (Ollama, LM Studio). Model performance varies — small models often have tool-use limits.

Workflow

The basic flow

From the CLI or local Web UI — all data lives in a local SQLite file.

📦

Install a skill

kittypaw skill install <name> — registry / GitHub URL / local path all supported.

⚙️

Configure

kittypaw skill config <name> <key> <value> — or kittypaw setup for LLM/channel in one shot. Required fields are in the package's package.toml.

⏱

Run / schedule

kittypaw skill run <name> for manual, or cron triggers driven by the server → output flows through the channel adapter.

Real scenarios

Flows built with registry packages

Live-tested ✅ vs in-progress 🚧 — see status on each card.

☀️

✅ Scheduled package execution

Install weather / market / RSS packages → cron schedule → telegram setup means delivery at the cron time.

🔄

✅ Repeat-intent surfacing

Reflection candidates stored → first-turn chat suffix verified (commit 8f2536a). Auto-generated code is 🚧.

🚨

✅ URL status check

cron HTTP check → alert message on status change. Telegram-routed delivery.

🌤

✅ Weather briefing

Open-Meteo 7-day forecast + LLM short summary, no API key needed.

💬

🚧 Draft a custom skill

kittypaw skill create "..." → JS code → goja syntax check (5/5 pass measured) → SkillStore save. Runtime stability depends on the package.

Registry

Packages currently in the registry

kittypaw skill install <name> to install. Status reflects live tests.

📈

Macro Economy Report

ETF + AI summary

✅ cron

🌤

Weather Briefing

7-day forecast (Open-Meteo)

✅ No API key

📰

RSS News Digest

Feed + LLM summary

✅ dupe filter

⏰

Reminder

Chat keyword + cron

🚧 partial

🔍

URL Monitor

HTTP check + status alert

✅ every 5 min

Tech stack

Pure Go — zero external runtimes

Single binary. No npm, Node, or Python. 20 SQLite migrations and 159 commits in.

🐹

Single binary · High performance

🖥

Web UI

Chi + WebSocket streaming

🔐

goja Sandbox

Safe skill execution

🤖

Claude + OpenAI

Multi-LLM support

🗄

SQLite

Local data storage

🔑

per-account secrets.json

Per-account isolated storage

🛡

Permission System

Approval-based access control

📦

TOML Package

Skill distribution format

KittyPaw A local AI automation experiment