MazenElhosseiny

Full Stack Chat Application With Multi Platform Clients AI Agent Integration — AI skill for Claude Code

AI community

A Claude Code-style AI agent harness in Python.

How to install Full Stack Chat Application With Multi Platform Clients AI Agent Integration

This entry records only its repository, not the path inside it, so there is no exact command to give. Open MazenElhosseiny/Full-Stack-Chat-Application-with-Multi-Platform-Clients-AI-Agent-Integration and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Full Stack Chat Application With Multi Platform Clients AI Agent Integration does

A Claude Code-style AI agent harness in Python. Uses Ollama + native LLM tool-calling to read/write files and execute shell commands in a prompt→execute→feedback loop. Sandboxed via Podman containers, with a self-modifying mode that lets the agent extend its own tools.

Alternatives in AI

Firecrawl MCP Server — 🔥 Official Firecrawl MCP Server - Adds powerful web scraping and search to Cursor, Claude and any other LLM c 6.1k ★
Sentrux — Real-time architectural sensor that helps AI agents close the feedback loop, enabling recursive self-improveme 2.9k ★
Multi Execute — 执行 - 多模型协作执行 (Multi-Model Collaborative Execution) 1.9k ★

README

AI Agent Harness

A Claude Code–style AI agent built from scratch in Python. It accepts prompts over HTTP, calls a local LLM through [Ollama](https://ollama.com) using the `openai` Python library, and executes tool calls (file I/O, shell commands) in a sandboxed container — including a self-modifying mode where the agent can extend its own toolset at runtime.

What it does

Receives a prompt via POST /chat
Sends the prompt + tool schemas to a local LLM (llama3.2:3b) using native tool calling
If the model requests a tool call, executes it and feeds the result back to the model
Repeats until the model returns a final answer (or hits a max-iteration safety limit)
Returns the final response as JSON

This is the same fundamental architecture behind tools like Claude Code and GitHub Copilot's agent modes: a **prompt → execute → feedback loop**, driven by the model's own decisions about when and how to use tools.

Tools implemented

Tool	Description
`read_file(path)`	Reads and returns the contents of a file
`write_file(path, content)`	Writes content to a file, creating parent directories as needed
`run_command(command)`	Executes a shell command and returns stdout/stderr

Each tool always returns a string — including on failure — so the model can see and reason about errors rather than the process crashing.

Three stages of progression

**Stage 1 — Bare process.** Run directly on the host. Demonstrates that without a sandbox, an agent with these tools has full access to the machine — nothing technical stops a destructive command from succeeding.

**Stage 2 — Sandboxed.** Run inside a Podman container. The agent can execute the exact same commands, but the container's filesystem namespace means destructive actions only affect the container's own isolated filesystem — the host is untouched.

**Stage 3 — Self-modifying.** With the source code mounted into the running container, the agent can rewrite its own `harne