Skip to content

Search RealDealCPA builds

Jump to any build, page or link

ShippedDeveloper tools

IronCore

The coding agent that molds itself to local models.

Coding agents like Claude Code and Codex shine with frontier models but fall apart on a smaller open model you run yourself. IronCore flips the assumption: it first measures what your model can actually do, then moves the jobs small models are bad at — remembering goals, formatting tool calls, applying patches, knowing when to stop — into reliable code.

View source on GitHub Install from PyPI
Offline tests
2,740
CI coverage gate
90%
Lines of Python
30.8k
GitHub releases
6
pypi.org/project/ironcore-cli/
IronCore terminal session with list_dir, read_file and grep tool cards

IronCore terminal session with list_dir, read_file and grep tool cards

Two ways to read it

In plain English

Coding agents like Claude Code and Codex shine with frontier models but fall apart on a smaller open model you run yourself. IronCore flips the assumption: it first measures what your model can actually do, then moves the jobs small models are bad at — remembering goals, formatting tool calls, applying patches, knowing when to stop — into reliable code.

For engineers

A Python 3.11+ Textual TUI and headless CLI with a JSON event stream. A background probe battery measures honest context, instruction retention, tool-call and edit-format reliability; the harness walks downgrade ladders (native function calls → guided-decoding strict JSON → a text protocol; unified diff → search/replace → whole file). Four safety modes, a policy engine, path jail, secret redaction, prompt-injection guards and git-snapshot undo.

What it does.

  • Measures each model's honest context and tool-call reliability before trusting it
  • Downgrade ladders guarantee a tool-call and edit format the model can produce
  • Four safety modes cycled with Shift+Tab; network access is never auto-allowed
  • Secret redaction, path jail, prompt-injection guards and git-snapshot undo
  • An “is it really done?” loop that refuses to report unverified work
  • Headless `ironcore exec` for scripts and CI, plus MCP servers, skills and plugins

The story

Advertised context is a lie. Measure it.

A 30B model that advertises 262,144 tokens of context may only stay honest for 49,152. IronCore finds that out in its first session, writes it into a model profile, and plans every turn around the real number — not the marketing one.

Built with

  • Python 3.11–3.13
  • Textual
  • Pydantic
  • uv
  • Ollama
  • llama.cpp
  • vLLM
  • MCP

Status: Shipped · started July 2026 · open on GitHub

Keep exploring