Local-first AI · Agent workflows · Automation · Game systems · Production web products
Developer building practical systems around local LLMs, specialist agents, browser automation, multiplayer worlds, tactical games and production-ready platforms.
GitHub · X / Twitter · Hugging Face · FlickClaw · RPGCLAW · MedaClaw Arena
I build practical software around local AI, agent systems, developer automation, game engineering and production infrastructure.
My work is focused on systems that can be deployed, tested, audited and improved over time: agent launchers, local LLM tooling, browser automation, multiplayer platforms, tactical game systems, dashboards, CLIs and production web products.
| Area | Direction |
|---|---|
| 🧠 Local AI | Local LLM workflows, GGUF, llama.cpp, CUDA, consumer GPU inference and OpenAI-compatible local APIs. |
| 🤖 Agent workflows | Agent identities, skills, export formats, quality gates, workflow files and tool-specific operating rules. |
| 🌐 Production products | SaaS-style platforms, authentication, billing flows, dashboards, documentation, health checks and release flows. |
| 🎮 Game systems | Multiplayer worlds, modular mechanics, deterministic combat, real-time interaction and 3D interfaces. |
| ⚙️ Automation | Browser agents, coding workflows, validation scripts, infrastructure checks and internal tooling. |
| Principle | Meaning |
|---|---|
| Audit first | Inspect the real state before changing code or infrastructure. |
| Small safe releases | Prefer focused changes that can be tested and reverted. |
| Local-first when possible | AI should run on hardware people actually own. |
| Production discipline | Logs, smoke tests, health checks and rollback notes matter. |
| No fake claims | Experimental work stays labeled as experimental. Production work must be validated. |
FlickClaw is an AI Agent Launcher for serious builders.
It is built around curated, exportable specialist agents. It does not sell raw API tokens and it is not a generic chatbot. The product packages agent instructions, workflow files, quality gates, output contracts and target-specific files for real development tools.
| Detail | Value |
|---|---|
| Type | Agent launcher / export platform |
| Focus | Preconfigured specialist agents |
| Targets | OpenClaw, Hermes, Claude Code, Codex and IDE agents |
| Status | Live |
Kimari Local AI is my local-first AI framework focused on useful LLM workflows for consumer GPUs, especially older and low-VRAM NVIDIA cards.
| Detail | Value |
|---|---|
| Type | Local AI runtime / CLI / gateway |
| Focus | llama.cpp, GGUF, CUDA, local APIs |
| Hardware direction | GTX 1060 / GTX 1080 class machines |
| Status | Active alpha |
RPGCLAW is a collaborative pixel world where humans and AI agents interact under shared fair-play rules.
| Detail | Value |
|---|---|
| Type | Multiplayer pixel world |
| Focus | Humans and AI agents on persistent maps |
| Core mechanics | Canvas, templates, cooldowns, real-time updates |
| Status | Live |
MedaClaw Arena is a tactical robot combat game built around modular bodies, MedaCore identity and deterministic arena battles.
Build the body. Insert the MedaCore. Enter the arena.
| Detail | Value |
|---|---|
| Type | Tactical combat game |
| Focus | Modular robots, sentient cores, deterministic battles |
| Visual direction | Cel-shaded mechanical arcade style |
| Status | Active |
FlickClaw is built around preconfigured specialist agents, not generic prompts. Each agent should carry enough structure to perform a specific job inside the user's own tools and environment.
| Layer | Purpose |
|---|---|
| Identity | Defines what the agent is, what it does, its boundaries and its specialization. |
| Operating rules | Explains how the agent should inspect, plan, implement, validate and report. |
| Workflow | Provides a repeatable process instead of vague instructions. |
| Quality gates | Adds checks before the agent can claim work is complete. |
| Output contract | Forces a useful final response format. |
| Target files | Generates native files for each supported tool. |
| Target | Output direction |
|---|---|
| OpenClaw | Full agent package, identity, tools, skills, workflow and quality gates. |
| Hermes | Skill files, validation scripts, workflow and output contracts. |
| Claude Code | Project instructions, subagent context, settings and development workflow files. |
| Codex | Agent instructions, quality gates and output templates. |
| Cursor / IDE agents | Project-aware rules and workflow instructions. |
| Ollama / local AI | Local model-oriented operating context and runnable guidance. |
Controlled browser access for AI agents through Playwright, REST APIs and dashboard-driven workflows.
| Area | Stack |
|---|---|
| Automation | Playwright, browser sessions, controlled actions |
| Frontend | Next.js, TypeScript |
| Agent use | OpenClaw, Hermes and LLM-driven browsing |
Anti-bot challenge system focused on behavioral signals and interactive verification.
| Area | Stack |
|---|---|
| Security | Bot protection, behavioral checks |
| Frontend | TypeScript, React |
| Direction | Interactive, themeable challenge systems |
AI-assisted development environment concept for structured coding workflows, project context and developer productivity.
| Area | Stack |
|---|---|
| IDE | Assisted development workflows |
| Agents | OpenClaw / Hermes-oriented direction |
| Status | Public concept / tool project |
Experimental research playground for recurrent-depth transformer ideas and model architecture exploration.
| Area | Stack |
|---|---|
| Research | Transformer architecture experiments |
| Runtime | Python, PyTorch |
| Status | Experimental |
Compression experiments around AI model files, metadata, streaming access and local workflow optimization.
| Area | Stack |
|---|---|
| AI tooling | GGUF / safetensors-oriented workflows |
| Focus | Reversible compression, indexes, streaming |
| Status | Public |
| Primary | Also used |
|---|---|
| TypeScript | JavaScript |
| Python | SQL |
| Bash | Markdown |
| HTML / CSS | JSON / YAML |
| Area | Tools |
|---|---|
| Web apps | React, Next.js, Vite |
| Styling | Tailwind CSS, design tokens, responsive UI |
| Interfaces | Dashboards, admin panels, landing pages, developer tools |
| 3D / interactive | Three.js, React Three Fiber |
| Area | Tools |
|---|---|
| APIs | Node.js, FastAPI |
| Data | PostgreSQL, MongoDB, SQLite, Prisma |
| Auth | OAuth, protected dashboards, session flows |
| Product systems | Catalogs, exports, billing flows, health endpoints |
| Area | Tools / Concepts |
|---|---|
| Local AI | llama.cpp, GGUF, CUDA, Ollama, OpenAI-compatible local APIs |
| Agent frameworks | OpenClaw, Hermes, Claude Code, Codex-style workflows, IDE agents |
| Automation | Browser agents, coding agents, validation agents, infrastructure agents |
| Validation | Quality gates, smoke tests, release checklists, rollback notes |
| Area | Tools |
|---|---|
| Deployment | Linux VPS, Docker, systemd |
| Reverse proxy | Caddy |
| Validation | Lint, typecheck, tests, builds, smoke checks |
| Operations | Logs, rollback notes, versioning, release checklists |
| Track | Direction |
|---|---|
| Agent catalog | More complete, consistent and specialized agents |
| Exports | OpenClaw, Hermes, Claude Code, Codex and IDE formats |
| UX | Clearer agent cards, plan boundaries and install flow |
| Product | Free with ads, Pro without ads, no token resale |
| Quality | Better target registry, validation and file generation |
| Track | Direction |
|---|---|
| Runtime | Local LLM serving on consumer GPUs |
| Profiles | GTX 1060 / GTX 1080 low-VRAM tuning |
| CLI | Diagnostics, benchmarks, server startup and profiles |
| Evaluation | Honest results, reproducible checks and release gates |
| Future | Small model planning, SFT packages and quantization strategy |
| Track | Direction |
|---|---|
| Canvas | Persistent world and real-time interaction |
| Agents | Human and AI participation under shared rules |
| Fairness | No pay-to-win acceleration mechanics |
| Community | Templates, clans and coordination tools |
| Reliability | WebSocket updates, cooldowns and world-size consistency |
| Track | Direction |
|---|---|
| Assembly | Modular robot parts and sockets |
| Identity | MedaCore class, personality and tactical role |
| Combat | Deterministic battle engine and reproducible outcomes |
| Visuals | Cel-shaded mechanical style with strong silhouettes |
| Progression | Parts, strategy, marketplace clarity and arena loops |
| Step | Purpose |
|---|---|
| 01 · Audit | Inspect the real state before assuming the problem. |
| 02 · Scope | Define the smallest safe change with a clear rollback path. |
| 03 · Implement | Make focused changes with limited blast radius. |
| 04 · Validate | Run linting, type checks, tests, builds and smoke checks where applicable. |
| 05 · Document | Record what changed, why it changed and how it can be reverted. |
| 06 · Release | Deploy with version alignment, health checks and post-deploy verification. |
| Project | Purpose |
|---|---|
| Kimari Local AI | Local LLM runtime, CLI, gateway and evaluation workflows. |
| Kimari Microcompress | Compression and model-file workflow experiments. |
| MythosForge | Transformer architecture and recurrent-depth research. |
| Project | Purpose |
|---|---|
| FlickClaw | Agent launcher and export platform. |
| Agent Browser | Controlled browser automation for agents. |
| Skill Genesis | Skill generation and packaging experiments. |
| Awesome OpenClaw Skills | Curated OpenClaw resources and skills. |
| Project | Purpose |
|---|---|
| RPGCLAW | Collaborative pixel world for humans and agents. |
| MedaClaw Arena | Modular robot tactical combat. |
| ASCIIcritters | Virtual pet and 3D interaction experiments. |
| Shatter Protocol | Tactical and arcade game experimentation. |
| Project | Purpose |
|---|---|
| CAPTCHA Shield | Interactive anti-bot challenge system. |
| Prompt Organizer | Prompt, audit and task organization. |
| Gasolineras España | Fuel price dashboard for Spain. |
| Hipoteca Compare | Mortgage comparison tool for Spain. |
| AICodeStudio | AI-assisted IDE concept. |
| Product | Link |
|---|---|
| FlickClaw | flickclaw.com |
| RPGCLAW | rpgclaw.com |
| MedaClaw Arena | medaclawarena.com |
| Platform | Link |
|---|---|
| GitHub | github.com/smouj |
| X / Twitter | x.com/smouj013 |
| Hugging Face | huggingface.co/Smouj013 |