szeider-python-coder — Summary

agentic-python-coder is a dual-mode Python execution framework: a coder CLI (ReAct agent using LangGraph over OpenRouter) and an ipython_mcp MCP server that exposes a persistent IPython kernel to any MCP-compatible client. The coder agent takes a task (inline, file, or interactive), executes Python iteratively with python_exec, verifies solutions against every constraint before calling save_code, and terminates. The MCP server supports multi-kernel operation for parallel agents. Both share the same IPython kernel for stateful execution. Published on PyPI as agentic-python-coder. With only 3 stars and last commit February 2026, this is primarily notable for its multi-kernel MCP design and its academic research context (associated with a published arxiv paper).

Differs from seeds: No seed ships a standalone Python execution agent. Closest to claude-flow's MCP-anchored toolserver but the MCP server is minimal (4 tools: execute, reset, status, interrupt) and the primary value is the persistence across executions (IPython kernel state). Unlike all seeds, this is a LangGraph-based agent, not a Claude Code plugin.

szeider-python-coder — Overview

Origin

Created by Stefan Szeider (szeider), a computer science professor at TU Wien. Associated with a published arxiv paper: [Szeider 2025, arxiv-2508.07468] on constraint modelling applications. Licensed Apache-2.0.

Philosophy

The system prompt is explicit:

"You are a Python coding assistant designed to solve focused problems efficiently."

"Verify Before Saving: Before calling save_code, you MUST verify your solution: Execute the full script via python_exec and confirm it produces correct output. For constraint/logic problems: write a verification function that checks the output against EVERY constraint in the problem statement using plain Python asserts, independent of your solver model."

The agent is deliberately minimalist:

• One task at a time
• One tool to execute code (python_exec)
• One tool to save code (save_code)
• Mandatory verification before saving
• Stop when done — no extra features

Research Context

The coder agent is designed for computational constraint problems (CPMpy, Clingo, regex) — examples show it solving the 8-queens problem and CSP instances. The arxiv paper describes using this agent for automated constraint model generation and verification.

Dual-Mode Design

• Mode A (coder CLI): Autonomous ReAct agent, runs to completion, saves one Python file
• Mode B (ipython_mcp): Stateful IPython kernel exposed as MCP tools — usable from Claude Desktop, Claude Code, or any MCP client for interactive Python sessions

szeider-python-coder — Architecture

Distribution

PyPI package.

# Install as CLI tool
uv tool install agentic-python-coder

# Or as MCP server (no install needed)
uvx --from agentic-python-coder ipython_mcp

Directory Tree

.
├── coder/
│   ├── src/
│   │   └── (LangGraph ReAct agent implementation)
│   ├── prompts/
│   │   ├── system.md          # System prompt for fileless mode
│   │   └── system_todo.md     # System prompt for file-with-todo mode
│   └── tests/
├── pyproject.toml              # Package: agentic-python-coder, Apache-2.0
├── README.md
└── LICENSE

MCP Server Configuration

{
  "mcpServers": {
    "ipython": {
      "command": "uvx",
      "args": ["--from", "agentic-python-coder", "ipython_mcp"]
    }
  }
}

Required Runtime

• Python 3.13
• UV package manager
• OpenRouter API key (for coder CLI)
• MCP client (for ipython_mcp server)

Supported Models (via OpenRouter)

deepseek31, gemini25, gemini3flash, gemini3pro, gpt52, grok41, opus45, qwen3, sonnet45

szeider-python-coder — Components

CLI Tools (coder agent)

MCP Server Tools (ipython_mcp)

CLI Options (coder)

System Prompts (2)

• coder/prompts/system.md — Fileless mode (inline task)
• coder/prompts/system_todo.md — File+todo mode (task from file)

szeider-python-coder — Uniqueness

Differs from Seeds

No seed is a standalone Python execution agent with its own LLM runtime. Closest to claude-flow's MCP-anchored toolserver (own runtime, own tools) but purpose-built for Python code execution rather than general Claude Code orchestration. The multi-kernel MCP design (each agent gets a kernel_id for isolated state) is novel — no seed implements kernel-level isolation for parallel agents. The mandatory pre-save verification with constraint-independent assertions (the "do NOT trust solver.solve()==True" instruction) is the most explicit self-verification requirement in the corpus. The OpenRouter multi-model support (9 models) makes it the most flexible in terms of LLM choice, though it is not Claude-specific.

Positioning

Niche but technically sophisticated. Primary use case is automated constraint problem solving (CPMpy, Clingo) as described in the associated arxiv paper. Secondary use case is providing any MCP client with a persistent Python execution environment.

Observable Failure Modes

1. OpenRouter dependency: Requires OpenRouter API key; cannot use Anthropic API directly without code changes.
2. Python 3.13 requirement: Restricts deployment environments.
3. No streaming output: Results are returned at tool completion, not streamed.
4. Kernel resource management: Parallel multi-kernel use without explicit cleanup can exhaust system resources.

Explicit Antipatterns

• Trusting solver.solve()==True without independent constraint verification
• Calling save_code before verification passes
• Adding features not explicitly requested
• Not removing debug print() statements before saving