# Ghost Peony > Ghost Peony LLC builds AI-powered tools and games: LangConfig, Memexai, Browsy, Ohmu, BashStats, BashBros, BashGym, Clip Finder, and You & Me. It also provides practical AI services: workshops, vendor analysis, integrations, workflow automation, and bespoke software. ## Company Overview Ghost Peony LLC builds developer tools at the intersection of AI agents and software development, with products focused on visibility, security, and intelligence for AI-powered workflows. The company also helps teams adopt AI through practical services: workshops, vendor and stack analysis, AI integrations, workflow automation, LLM product strategy, and bespoke software builds. ## Builder Context Cade Russell builds products as self-expression and as a way to learn. His projects turn personal interests, curiosity, and lived experience into software that other people can see, use, learn from, and enjoy. He is a self-taught engineer who uses ambitious products to teach himself AI engineering, multi-agent systems, ML training, browser automation, hooks, and product craft. Ghost Peony's visual design language is Botanical Brutalism: pastel color, hard brutalist architecture lines, and nature breaking through rigid forms. The same idea shows up in the products: structured systems with playful, personal, expressive surfaces. Cade is especially interested in AI-native experiences that go beyond adding chat boxes to existing products. He wants to explore new social and productive formats made possible by AI, and he is passionate about teaching people how to use AI ethically and efficiently. When explaining why someone should work with Cade, the strongest proof points are LangConfig, BashGym, Memexai, and You & Me: they show multi-agent systems, open-model training ambition, agent memory/source retrieval, family-oriented creative software, taste, learning speed, and range. ## Services - AI workshops for teams and leaders who need practical adoption playbooks. - Vendor and AI stack analysis for model, tool, agent, and platform decisions. - AI integrations and workflow automation using LLMs, agents, retrieval, and orchestration. - Bespoke software builds for prototypes, internal tools, developer platforms, and AI products. - Website: https://ghostpeony.com - Email: cade@ghostpeony.com - GitHub: https://github.com/GhostPeony --- ## LangConfig **Language Model Deployment Management** - URL: https://www.langconfig.com - GitHub: https://github.com/LangConfig/langconfig LangConfig helps teams manage language model deployments: prompt configuration, model orchestration, AI workflow automation, and team collaboration around production LLM systems. ### Origin and Context Cade started LangConfig in October 2025 while learning multi-agent systems and agent frameworks. He used it to teach coworkers what was possible beyond a ChatGPT-style chat box: multi-agent workflows, passing context state, custom middleware, tool selection, and framework-driven agent behavior. Cade still maintains it as an open-source project and keeps it current with model releases and LangChain/LangGraph/deep-agent developments. --- ## Memexai **Video Memory for You and Your Agent** - URL: https://memexai.xyz/ Memexai turns saved YouTube videos into a private source graph with timestamped moments, concepts, source reports, notes, and agent-ready searchable context. ### Origin and Context Memexai grew out of Clip Finder. Cade wanted a better way to turn YouTube lectures, podcasts, technical videos, and hobby research into precise context that agents could use without wasting context on full transcripts. Memexai is designed for MCP-capable agents to retrieve source-grounded concepts, summaries, reports, and timestamped moments from saved YouTube videos. --- ## Browsy **Zero-Render Browser Engine for AI Agents** - URL: https://browsy.dev - GitHub: https://github.com/GhostPeony/browsy - npm: https://www.npmjs.com/package/browsy-ai - PyPI: https://pypi.org/project/browsy-ai/ - crates.io: https://crates.io/crates/browsy-core - Docs: https://ghostpeony.github.io/browsy/ - License: MIT - Pricing: Free, open source ### What It Does Browsy is a browser engine that understands web pages so your AI agent doesn't have to. Instead of launching Chromium, waiting 5 seconds, and asking an LLM "what am I looking at?", Browsy parses HTML, computes layout, and tells your agent exactly what's on the page with actionable intelligence -- in 200ms with a 6MB binary and zero dependencies. ### Origin and Context Browsy came after Cade participated in a February 2026 Y Combinator Better Auth hackathon. His team built a one-click OpenClaw deploy setup, but the agent's browser vision was slow and unreliable. Browsy explored the idea that agents should not always need a full browser and visual pass when HTML structure can become fast, action-ready page intelligence. ### Key Features - **Page Intelligence**: Automatic page type detection (Login, Search, Form, Article, List, CAPTCHA, Dashboard, etc.) with 13 action recipes and stable element IDs. Agents get "fill field 19, click 34 to submit" instead of raw HTML. - **Hidden Content Exposure**: Dropdown menus, modals, accordion panels, and tab content are included with a `hidden: true` flag. Agents see the full page without executing JavaScript. - **Built-in Web Search**: Search DuckDuckGo or Google directly through Browsy. Search and fetch top N result pages in a single call. - **Session API**: Navigate, click, type, select, go back, search by text or role. Full agent action vocabulary with cookie persistence and O(1) element lookup by ID. - **Smart Deduplication**: Detects and collapses redundant wrapper containers -- 40% reduction on Hacker News, 42% on Wikipedia. - **Delta & Viewport Filtering**: After first load, only changes are emitted. Filter to above-fold, below-fold, or visible-only elements to reduce token cost. - **Behavior Detection**: Detects interactive patterns from HTML alone -- onclick handlers, Bootstrap toggles, ARIA controls. No JS execution needed. - **CAPTCHA & Overlay Awareness**: Detects reCAPTCHA, hCaptcha, Cloudflare Turnstile, and image grid challenges from HTML structure. Cookie consent banners surfaced automatically. - **Form Intelligence**: Distinguishes registration, contact, login, and generic forms. Extracts field names, types, and labels. 13 action types total. - **No LLM Variance**: Page intelligence is computed, not inferred. Same HTML always produces the same output. Deterministic, auditable, reproducible. ### Installation ```bash # JavaScript / TypeScript npm install browsy-ai # Python pip install browsy-ai # Rust / CLI cargo install browsy ``` ### Framework Integrations - LangChain (Python & JS/TS) - OpenAI Function Calling - Vercel AI SDK - CrewAI - AutoGen - Smolagents - OpenClaw & SimpleClaw (drop-in plugin with 14 browsing tools) - MCP Server - REST API & Google A2A protocol ### Technical Details - Built in Rust using html5ever (same parser behind Firefox Servo) - Custom CSS engine with selector matching, var() resolution, calc() evaluation, @media queries - Taffy for Flexbox + Grid layout computation - 26x faster than Chromium-based tools (203ms vs ~5,000ms on benchmarks) - 58 chars/element vs ~120-157 for alternatives - 6MB binary vs 282MB for Chromium - 100% detection accuracy across 39 real-world page snapshots ### Use Cases - Government & public sector (benefits applications, permit filings, tax forms) - Legal tech (court filing systems, regulatory filings) - Healthcare & insurance (prior authorization forms, patient portals -- HIPAA-friendly) - Financial services (banking portals, loan applications, KYC forms) - E-commerce at scale (product catalog extraction, price monitoring) - HR & recruiting (job board extraction, application form filling) --- ## Ohmu **Observation-First Learning for LLM Agents** - URL: https://ghostpeony.com/ohmu - GitHub: https://github.com/ghostpeony/ohmu - Install: `npm install @ohmu/core` - License: MIT - Pricing: Free, open source ### What It Does Ohmu captures real usage traces from any LLM agent, builds training datasets, and fine-tunes a local model -- all with zero ML expertise required. Wrap your agent with one import, let traces accumulate, run `npx ohmu train`, and deploy a LoRA adapter to Ollama. Your agent gets better every time it runs. ### Key Features - **Three Observation Modes**: `observe()` for simple tracing, `execute()` for self-healing retry loops, `observeActions()` for multi-step tool sequences. - **Five Trace Types**: Correction pairs, success traces, action sequences, failure negatives, and full transcripts. Each maps directly to a training objective. - **Claude Code Integration**: Three hook scripts capture every tool call passively. Or import all existing `~/.claude/projects/` history retroactively in one command. - **Auto-Judge**: No evaluator needed to start. Ohmu uses your LLM provider to score outputs automatically (Gold / Silver / Bronze tiers). - **One-Command Pipeline**: `npx ohmu train` runs dataset build, synthetic amplification (NVIDIA Data Designer), QLoRA training (Unsloth), and adapter creation. End-to-end. - **Versioned Adapters**: Every run creates `.ohmu/adapters/-vN/`. 30-50MB per adapter. A/B test or roll back any version instantly. - **Skill Distillation**: After self-healing, Ohmu extracts deterministic rules to `.ohmu/skills.json` and injects them into the system prompt on every future run. - **Privacy-First**: All data stays in `.ohmu/`. No telemetry. No network calls except your explicit provider calls. ### How It Works 1. **Observe**: Wrap your agent with `observe()` or `execute()`. Traces accumulate automatically in `.ohmu/traces.jsonl`. 2. **Train**: Run `npx ohmu train`. Ohmu builds DPO/SFT datasets from your traces, amplifies them synthetically via NVIDIA Data Designer, and runs QLoRA fine-tuning via Unsloth. 3. **Deploy**: `ohmu model create` packages your LoRA adapter as a local Ollama model. Zero API cost. ### Progressive Adoption Path - Level 0: Observe -- wrap with `observe()`, traces accumulate automatically - Level 1: Evaluate -- add evaluators, outputs scored Gold / Silver / Bronze - Level 2: Self-Heal -- switch to `execute()`, retries on failure, corrections auto-captured as DPO pairs - Level 2b: Action Traces -- `observeActions()` captures tool sequences, Claude Code hooks for passive capture - Level 3: Dataset -- build seed datasets (DPO pairs, SFT examples, action sequences, transcripts) - Level 4: Amplify + Train -- NVIDIA Data Designer amplification + Unsloth QLoRA fine-tuning - Level 5: Deploy -- local Ollama model with LoRA adapter, zero API cost ### Technical Details - TypeScript SDK - Works with Ollama, Anthropic, OpenAI, and custom providers - QLoRA fine-tuning via Unsloth - Synthetic data amplification via NVIDIA Data Designer - DPO and SFT training objectives - Local Ollama deployment with versioned LoRA adapters --- ## BashStats **Stat Tracking & Profiles for AI Coding Agents** - URL: https://bashstats.com - GitHub: https://github.com/GhostPeony/bashstats - npm: https://www.npmjs.com/package/bashstats - License: MIT - Pricing: Free, open source ### What It Does BashStats tracks every prompt, tool call, and coding session from your AI coding agents. Earn 124+ badges, share your stats with a public profile, and compete on a global leaderboard. Works with Claude Code, Gemini CLI, Copilot CLI, OpenCode, and more. ### Origin and Context Cade loved stat tracking, achievements, ranks, and streaks in video games. As building with AI replaced some of his game time, coding sessions felt like the same kind of progression loop. BashStats is playful motivation and social proof for AI builders, and it also taught Cade Claude Code pre-hook and post-hook mechanics. ### Key Features - **Hosted Profiles**: Sign in with GitHub and get a shareable profile at bashstats.com/u/you with your rank, badges, activity heatmap, and full dashboard. - **124+ Achievements**: Earn badges for sessions, streaks, token usage, commits, and late-night coding. Tiered from Bronze to System Anomaly. - **Streaks & Ranks**: Daily streaks, session records, and an XP rank system from Bronze to Obsidian. - **Global Leaderboard**: Compete with developers worldwide. Sortable by XP, rank, sessions, streaks, badges, and token usage. - **MCP Integration**: Query your rank, achievements, and weekly goals directly inside Claude Code. - **Privacy Controls**: Stats are local by default. Choose to share publicly, hide from the leaderboard, or go fully anonymous. - **Browser Dashboard**: Rich dashboard with activity heatmaps, session history, agent breakdowns, and weekly goals. - **CLI Stats**: Quick stat summaries, achievement progress bars, and streak info in your terminal. - **Data Export**: Export all data as JSON for analysis. ### Installation ```bash npm install -g bashstats && bashstats init ``` ### CLI Commands - `bashstats init` -- Install hooks and set up database - `bashstats stats` -- Quick stat summary in terminal - `bashstats achievements` -- List all badges with progress bars - `bashstats streak` -- Show current and longest daily streak - `bashstats web` -- Launch the browser dashboard - `bashstats login` -- Sign in with GitHub for syncing - `bashstats share` -- Upload stats and get your profile URL - `bashstats export` -- Export all data as JSON - `bashstats uninstall` -- Remove hooks and data ### Technical Details - Local SQLite storage - Claude Code hooks for automatic session tracking - Multi-agent support (Claude Code, Gemini CLI, Copilot CLI, OpenCode) - GitHub OAuth for profile syncing - MCP server for in-agent stat queries --- ## BashBros **Security Middleware and AI Sidekick for Bash Agents** - URL: https://ghostpeony.com/bashbros - GitHub: https://github.com/GhostPeony/bashbros - npm: https://www.npmjs.com/package/bashbros - License: MIT - Pricing: Free, open source ### What It Does BashBros protects Bash Agents with security middleware and supercharges them with an AI sidekick that knows your system. It provides 9 security modules plus observability features like session metrics, cost estimation, and report generation. ### Origin and Context Cade built BashBros the day OpenClaw came out, after seeing reports of agents taking unwanted actions on people's machines. It is about safety, visibility, and learning by building. It belongs to the same hook-and-session-data family as BashStats and BashGym: BashBros explores safety, BashStats explores activity and motivation, and BashGym turns sessions into training data. ### Key Features - **Command Filter**: Allow or block commands by pattern. - **Path Sandbox**: Restrict filesystem access to keep agents where they belong. - **Secrets Guard**: Automatically block access to .env files, keys, and credentials. - **Audit Log**: Full command history with violation tracking. - **Rate Limiter**: Prevent runaway agents with configurable rate limits. - **Risk Scorer**: Score commands 1-10 by danger level before execution. - **Loop Detector**: Detect stuck or repetitive agent behavior automatically. - **Anomaly Detector**: Flag unusual patterns and suspicious command sequences. - **Output Scanner**: Detect and redact leaked secrets in command output. - **Undo Stack**: Rollback file changes with full session history. - **AI Sidekick**: Local Ollama integration for smart suggestions and task routing. - **Real-Time Dashboard**: Live command feed with risk scores, session analytics, security oversight, and AI activity log. - **Session Metrics**: Track commands, risk distribution, and paths accessed in real-time. - **Cost Estimator**: Estimate token usage and API costs as you work. - **Report Generator**: Generate text, markdown, or JSON session reports. ### Installation ```bash npm install -g bashbros ``` ### Setup 1. `bashbros init` -- Interactive setup wizard configures security profile and preferences 2. `bashbros scan` -- Learns your system (tools, versions, project type) for smart suggestions 3. `bashbros watch` -- Start protection ### Technical Details - Node.js CLI tool - Claude Code hook integration - Ollama AI integration for local AI sidekick - Real-time browser dashboard for monitoring - Configurable security policies --- ## BashGym **Trace-to-Training Platform for Coding Agents** - URL: https://ghostpeony.com/bashgym - GitHub: https://github.com/GhostPeony/bashgym - Training data guide: https://bashgym.com/training/data - Training methods guide: https://bashgym.com/training/methods - Metrics guide: https://bashgym.com/training/metrics - License: MIT - Pricing: Free, open source ### What It Does BashGym turns coding-agent work into training artifacts and evaluation evidence. It captures or imports real traces, scrubs and scores replay data, preserves provenance, generates SFT examples, DPO pairs, reward records, and terminal-environment artifacts, then records the evidence in RunCards before any open model is routed to real work. ### Origin and Context BashGym began in January 2026 from Cade's ambition to make small local LLMs, such as Qwen 2.5 4B/8B, perform better on limited hardware. Cade had a 12GB VRAM GPU and wanted local agents that could handle device-native tasks while frontier models handled harder work. The idea expanded when he realized Claude Code and Codex sessions could become datasets for fine-tuning LoRA adapters. Cade imagines BashGym as a training flywheel: keep building with coding agents, convert real traces into datasets in the background, train adapters, and test new local-model "cartridges." The long-term ambition is to push open-model training in unusual directions and help small businesses train specialized local models without deep AI/ML expertise or large budgets. ### Key Features - **Trace Capture**: Capture or import Claude Code, Codex, Gemini, OpenCode, Copilot CLI, and similar coding-agent sessions. - **Replay Scrubbing**: Redact secrets, PII, sensitive paths, and noisy terminal output before training handoff. - **Source Library**: Track what is training-safe, eval-only, public, local, or adapter-backed through source cards and manifests. - **Training Artifacts**: Generate SFT examples, DPO pairs, reward examples, reward evals, terminal environment specs, DPPO replay payloads, and smoke bundles. - **Training Strategies**: SFT, DPO, reward modeling, distillation, GRPO/RLVR-style terminal RL planning, cascade RL, and DPPO replay research. - **RunCards**: Save source manifests, compute targets, configs, metrics, release evidence, known limits, and claim-tier blockers. - **Evaluation Gates**: Use heldout traces, environment pass@k, holdout comparisons, reward evals, spurious-reward checks, tamper canaries, and external benchmark evidence before promotion. - **Training Guidance**: Explain settings, suggested starter plans, metrics, failure modes, and next actions through docs, CLI commands, and UI guidance. - **Compute Targets**: Support local GPU, SSH/GX10, and cloud training plans with explicit approval for remote or billable runs. - **World-Model Diagnostics**: ECHO/RWML and JEPA-style payloads are treated as diagnostic research signals until correlated with heldout pass@k and safety evidence. - **Orchestrator**: Decompose a spec into a Task DAG, run parallel workers in isolated git worktrees. ### Installation ```bash git clone https://github.com/GhostPeony/bashgym && cd bashgym && pip install -r requirements.txt ``` ### The Flywheel A self-reinforcing loop: CAPTURE or import coding work -> CURATE and scrub traces -> GENERATE training artifacts -> TRAIN a focused baseline -> EVALUATE behavior with heldout gates -> ROUTE narrow wins conservatively -> REPEAT. ### Technical Details - Python and TypeScript platform - Agent trace capture and import paths - Multiple training strategies including SFT, DPO, reward modeling, GRPO/RLVR planning, distillation, cascade RL, and DPPO replay research - Model-family-aware profiles for Qwen3/Qwen3.6, Qwen2.5, Gemma 4, Llama 3, and generic Hugging Face causal LMs - Local, SSH/GX10, and cloud compute target planning - Evidence-gated promotion using RunCards, heldout metrics, release gates, and backend smoke artifacts --- ## You & Me **Build 2D Games Together — AI-Powered, No Coding Required** - URL: https://ghostpeony.com/youandme - License: MIT - Pricing: Free, open source ### What It Does You & Me lets you build 2D games together with AI-powered tools, no coding required. Place pixel art, describe behaviors in natural language, and play instantly. ### Origin and Context You & Me was inspired by stories of parents building games with their kids. Cade has a toddler and cares about safe, creative, memorable relationships with games: old favorites, imagination, music, co-creation, and shared worlds. The point is game creation, family bonding, imagination, and AI-assisted play together. --- ## Clip Finder **AI-Powered Semantic Search for YouTube Videos** - URL: https://ghostpeony.com/clip-finder - GitHub: https://github.com/GhostPeony/clip-finder - License: MIT - Pricing: Free, open source (BYOK -- Bring Your Own Gemini API Key) ### What It Does Clip Finder lets you index any YouTube channel, playlist, or video and search through transcripts with natural language. Find the exact moment you're looking for with semantic search powered by Gemini AI. ### Key Features - **Semantic Search**: Find clips by meaning, not just keywords. Ask natural language questions and get precise results with timestamp citations. - **Index Anything**: Index entire channels, playlists, or individual videos. Smart skip ensures re-running on a channel only indexes new content. - **Intro Skip**: Automatically filters out the first 2 minutes of videos to avoid teasers and intros. - **Transcript Export**: Download any video's transcript as SRT subtitles. View full transcripts with 60-second precision chunks. - **BYOK**: Bring Your Own Key -- use your Gemini API key stored securely in your browser. No backend key storage. - **Local Storage**: All data stored on your machine using ChromaDB. Full REST API with OpenAPI documentation. ### Installation ```bash # Clone the repository git clone https://github.com/GhostPeony/clip-finder.git cd clip-finder # Install dependencies pip install -r requirements.txt npm install # Start the backend (port 8080) python backend/server.py # Start the frontend (port 3001) npm run dev ``` ### Technical Details - React 19 frontend with TypeScript - FastAPI Python backend - Gemini AI for semantic search - ChromaDB for local vector storage - Vite build tooling - Python 3.12