Pengy is a local-first AI agent that lives on your machine — not in a browser tab, not behind a corporate API wall, not tethered to someone else's data center. It's a desktop application (PySide6 GUI + slick CLI) that pairs an LLM with real tools that can read your files, run your code, probe your network, and act on your system.
This isn't a chatbot that answers questions. It's an agent that does things.
Forget copying error messages into ChatGPT and getting generic advice back. Pengy has root access (when you give it), runs bash commands, SSHes into your home lab, edits configs, and fixes things — all in a single conversation.
Real examples:
- Recompile llama.cpp with CUDA support on a remote GPU server, diagnose build failures (missing PATH), fix the script, re-run it
- Audit nftables firewall rules on a router, identify missing allow rules, add them live
- Build and deploy router backup scripts that snapshot WireGuard configs, nftables rules, DNS settings — then schedule them via systemd timers
- SSH into machines across the house to run updates, clean old kernels, and reboot — all from one window
- Diagnose a broken GRUB/Plymouth boot screen, recreate the missing /etc/default/grub from scratch, and run update-grub
- Upgrade 41 apt packages, clean orphans, refresh snaps, recommend a reboot — in under a minute
vs. a basic chatbot: You'd paste error text, get "try updating grub," go back and forth five times, and still end up doing all the typing yourself.
Chatbots give you answers. Pengy gives you capabilities that persist. When you need to do something repeatedly, you build a skill — a Python script + documentation that Pengy knows how to find and use.
Real skills created by real users: - Weather skill — fetches hyperlocal forecasts from tomorrow.io, handles rate limiting with exponential backoff, caches results - Plot skill — generates matplotlib charts (line, bar, scatter, pie, histogram) with dark theme, multi-series support - Image generation skill — creates AI art via Google Gemini, saves PNGs to your Pictures folder - Image editing skill — modifies existing images by text prompt (add a maple leaf to that penguin, put sunglasses on that corgi) - Sensors skill — reads temperature/humidity from 6 ESP8266/BME280 nodes scattered around the house, answers "What's the hottest room?" - YouTube transcript skill — fetches transcripts from any video URL; when YouTube blocks the IP, auto-falls back to youtubesummary.com - Reddit skill — fetches Reddit content through old.reddit.com to evade AI-blocking - Email skill — sends rich HTML emails via Gmail SMTP with multipart/alternative fallbacks - Clip skill — posts formatted content to a self-hosted text clipping service for shareable URLs - RSS/News skill — parses any RSS/Atom feed using only stdlib, maps topics to feeds for intelligent routing - User profile skill — stores personal context so Pengy knows who it's talking to - TTS skill — speaks text aloud on Ubuntu using the best available engine - Network skill — inventories every machine on the LAN with specs, IPs, GPUs, running services
Skills are composable. Pengy can chain them: fetch weather → generate a chart → email it → clip the summary. The whole pipeline in one request.
vs. a basic chatbot: You get an answer you have to manually act on. With Pengy, you build a tool once and it becomes a permanent part of the agent's capability set.
Open a project folder and ask Pengy to review it. It will explore the directory tree, read multiple files in parallel, search for patterns across the codebase, and give you real architectural feedback — not just surface-level commentary.
Real examples: - Audited the entire Pengy app (12 source files across core, CLI, UI) against its design spec, identifying gaps, bugs, and improvement opportunities - Analyzed openai-proxy codebase to determine if model names were immutable primary keys — then built a migration script to rename them across both models and usage records - Debugged a double-encoding bug in HTML entity rendering (… showing as literal text), traced it through the escaping pipeline, fixed it - Diagnosed a single-shot CLI bug where user prompts were sent before being appended to the message list — found the exact line, reordered operations, verified fix - Fixed 502/400 proxy errors caused by duplicate Authorization headers and empty API key handling - Analyzed 45 chat sessions across multiple machines to compile usage patterns and build this very marketing material
vs. a basic chatbot: You'd paste fragments of code, ask "what's wrong here," get vague suggestions, and spend hours chasing wrong leads. Pengy reads the entire codebase and uses real filesystem tools to explore.
Pengy knows your home network. It has a persistent inventory of every machine, its IP, specs, running services, and codec support. It SSHes into machines to gather info, apply configs, and manage services.
Real examples: - Discovered a server had migrated from DigitalOcean droplet to a fresh Ubuntu 24.04 server, then updated the network documentation with full specs, listening ports, and service inventory - Fixed a Postfix mail relay on a new droplet by backing up configs from the old server, transferring them, validating every file, and flagging missing SPF/DKIM/PTR records - Identified 5 machines that would benefit from Vulkan Video hardware decoding in Firefox 153 by cross-referencing bug reports against GPU inventory - Analyzed Docker containers, MongoDB versions, and nginx configs on a remote server - Set up passwordless sudo across the entire lab for remote management, documented the setup
vs. a basic chatbot: It doesn't know your infrastructure exists. Pengy maintains a living document of your network and acts on it.
Pengy doesn't just answer questions — it investigates, cross-references, and delivers structured analysis.
Real examples: - Researched Tesla Model 3 vs Hyundai Ioniq 5 N with detailed specs, pricing, and a personality-based recommendation - Compared AI model pricing across DeepSeek, OpenAI, and Anthropic — including prompt caching math to calculate real costs at Pat's usage level of ~9M tokens/month - Analyzed a 306-comment Hacker News thread about anti-AI sentiment, extracting both sides of the debate with specific arguments - Researched Mark Carney's trade deal success as PM, pulling from Guardian, Al Jazeera, CBC, and Globe and Mail - Compiled summer 2026 weather outlook for all of Canada from a 20-minute YouTube video transcript - Investigated GIC investment purchasing processes across Questrade and RBC Direct Investing, compiling step-by-step guides for both platforms - Investigated whether a 27-year-old medical intern's Reddit post was about depression or burnout, summarizing the full discussion and community response
vs. a basic chatbot: You get a single-sentence answer from a generic web search. Pengy fetches multiple sources, reads full articles, and synthesizes structured analysis.
Pengy generates images, edits them, creates charts, speaks text aloud, generates sound effects, and summarizes video content.
Real examples: - Generated a cartoon penguin in a top hat, then edited it to add a Canadian maple leaf — all by text prompt - Created weather comparison charts for 3 cities with matplotlib - Generated 7 WAV sound effects (laser blasts, explosions, sirens, bird chirps) via SoX - Wrote a TTS script that read a news script aloud - Generated anime-style corgi art and boxer dog pool portraits
vs. a basic chatbot: It describes what an image might look like. Pengy makes the image, edits it, and saves it to your Pictures folder.
Pengy composes and delivers personalized daily briefings via email — weather, news, deals, everything — without any human in the loop.
Real example flow: 1. Fetches local weather from tomorrow.io 2. Pulls top stories from CBC, BBC, Hacker News RSS feeds 3. Scans RedFlagDeals for hot Canadian deals 4. Reads temperature/humidity sensors throughout the house 5. Formats everything into a rich HTML email with gradients and emojis 6. Sends it to the user's inbox
All triggered by a single command: "Send me today's briefing."
vs. a basic chatbot: You can't automate anything. Every query requires manual intervention.
Pengy is built for people who care about control, privacy, and flexibility:
~/Personal/skills/The pattern is always the same: Python script + markdown docs = permanent capability.
| Basic Chatbot | Pengy | |
|---|---|---|
| Answers questions | ✅ Yes | ✅ Yes + sources them from your actual system |
| Runs code | ❌ No | ✅ bash, Python, SSH, systemd, whatever |
| Edits files | ❌ No | ✅ Reads, writes, edits, searches across your projects |
| Knows your network | ❌ No | ✅ Full inventory, SSH access, config management |
| Persistent skills | ❌ No | ✅ Build once, use forever |
| Local models | ❌ Cloud-only | ✅ Any OpenAI-compatible endpoint, including local |
| Data privacy | ❌ Cloud-dependent | ✅ Fully local option |
| Automation | ❌ Manual only | ✅ Scriptable, composable, scheduleable |
| Extensible | ❌ Closed | ✅ Open source, hackable, skill-based |
| File system access | ❌ None | ✅ Full read/write/search/explore |
pip install pengy-agent
Or build from source: github.com/beholder/Pengy
Local-first. Model-agnostic. Your system, your rules.