WrongStack/WrongStack

An AI coding agent that reads your code, edits files, runs commands, and reasons through bugs — across a terminal REPL, a full-screen TUI, and a browser UI, while you keep your hand on every permission.

What it solves

WrongStack is an autonomous AI coding agent designed to handle complex software engineering tasks without requiring a subscription or relying on a third-party orchestration layer. It solves the problem of "thin wrapper" AI tools by providing a complete, standalone engine that manages its own memory, tool execution, and multi-agent coordination, allowing it to become more familiar with a specific codebase over time.

How it works

WrongStack operates on a compact kernel that manages a pipeline of events and controllers. It uses a "Director" to fan out tasks to a roster of 77 specialized agent roles, which coordinate via a project-wide inter-agent mailbox. The system employs a project-local long-term memory system called SAGE (backed by SQLite/FTS5) that anchors knowledge to specific files, symbols, or commits, auto-injecting relevant context into the agent's prompts. It supports a wide array of LLM providers (including local ones like Ollama) and offers six different user interfaces, ranging from a simple REPL to a full Desktop app and a cross-machine HQ Command Center.

Who it’s for

It is built for developers who want a powerful, autonomous AI coding assistant that can run locally, maintain durable project memory, and integrate deeply with their existing codebase without being locked into a specific provider's ecosystem.

Highlights

  • Multi-Agent Fleet: A 77-role roster coordinated by a Director with isolated budgets and transcripts.
  • SAGE Long-Term Memory: Code-anchored, persistent memory that remembers decisions and conventions across sessions.
  • Chimera Auto-Review: Dedicated agents that critique diffs and provide one-line fixes.
  • Extensive Toolset: 67 built-in tools for filesystem edits, code quality checks, web search, and E2E controls.
  • Flexible Interfaces: Six surfaces including TUI, WebUI, Desktop, and a cross-machine HQ for team coordination.
  • Provider Agnostic: Support for ~140 providers, including local endpoints (Ollama, vLLM) and OAuth sign-in for existing subscriptions.

Related

  • Project
    TKasperczyk/thunderbird-mcp

    An MCP server that gives AI assistants full access to Mozilla Thunderbird, enabling them to search mail, compose messages, and manage calendars and contacts.

  • Project
    CGCL-codes/naturalcc

    A sequence modeling toolkit for code intelligence that enables training and evaluating models for tasks like code generation, summarization, and retrieval.

  • Project
    fabriziosalmi/UglyFeed

    An LLM-powered application that retrieves, aggregates, and rewrites RSS feeds to create curated, high-quality content streams.

  • Project
    iMoonLab/Hyper-RAG

    A hypergraph-driven Retrieval-Augmented Generation (RAG) framework that reduces LLM hallucinations by capturing complex, beyond-pairwise relationships in domain-specific knowledge.