coco-research/coco

CoCo Super Intelligence is the orchestration layer that turns Claude Code, Cursor, or Codex into an engineering department: a routed advisory board, 226 skills, 386 commands, persistent state. Local. Open-core — MIT core; Super Intelligence is proprietary, own-use.

What it solves

CoCo Super Intelligence provides an orchestration layer for AI coding assistants (like Claude Code, Cursor, or Codex) to transform a single-prompt experience into a coordinated engineering department. It solves the problem of fragile chat contexts and the lack of of structured, expert-led decision-making and verification in AI-driven development.

How it works

CoCo is a vendor-neutral layer of Markdown and YAML that installs into existing agent harnesses. It does not run its own model but instead upgrades the host tool with several key components:

  • The Board: A routed expert panel of 495 named personas across 13 departments (e.g., Engineering, AI, Finance, Legal) that can deliberate, debate, and provide attributed, grounded verdicts on architectural or business decisions.
  • Execution Framework: A pipeline that uses specialized agents (Research, Architecture, Plan, Review, Build, QA) to execute decisions through six development stages and seven verification gates.
  • Architecture Conformance: A system that uses a commit-pinned index (.arch/index.json) to detect structural drift and ensure the final build matches the declared architectural plan.
  • Persistent State: Saves phase state and progress logs to disk, allowing developers to resume long-running tasks across session resets.

Who it’s for

Software engineers and architects using AI coding tools who need high-level strategic guidance, rigorous verification gates, and a way to manage complex, multi-stage engineering projects without losing context.

Highlights

  • 495 Expert Personas: A massive library of specialized minds for cross-domain decision-making.
  • Architecture Gates: Prevents "architectural dissolution" by verifying that code lands exactly where the plan specified.
  • Librarian of Skills: Includes 226 pre-defined skills and 386 slash commands for specific engineering and product tasks.
  • Local-First: 100% local installation with no telemetry and no requirement for new AI models.

Related

  • Project
    carloslfu/slotstream

    slotstream is a Swift‑based inference engine that runs the 125‑billion‑parameter Qwen‑3.8‑Flash‑Next model on Apple‑silicon Macs. It streams 4‑bit‑quantised weights from SSD into a RAM cache, enabling ~12 tokens / s on a 48 GB Mac. The tool offers a CLI, Ollama‑compatible and OpenAI‑compatible HTTP APIs, optional vision support, speculative decoding, and a Swift package for embedding in apps.

  • Project
    RealAlexandreAI/json-repair

    A Go-based toolkit for automatically fixing malformed JSON output generated by LLMs, ensuring it becomes valid and parseable.

  • Project
    facebookincubator/dynolog

    Dynolog is a lightweight performance monitoring daemon for heterogeneous CPU-GPU systems, designed to help debug bottlenecks in large-scale AI training. It provides always-on monitoring of CPU, GPU, and kernel metrics, plus on-demand remote profiling of PyTorch workloads across many GPUs.

  • Project
    tensorflow/cloud

    A set of APIs that allow developers to easily move TensorFlow and Keras model training from local environments to distributed training and tuning on Google Cloud Platform.