Ryder-Sun/Meldwork

Local-first AI agent workspace for multi-agent collaboration, agent orchestration, scoped permissions, evidence-aware runs, and human-in-the-loop decisions.

What it solves

Meldwork eliminates the need to manually copy context and prompts between multiple terminal windows when using different AI coding agents. It provides a unified desktop workspace to coordinate various local agent CLIs, allowing users to compare their outputs side-by-side and ensure no changes are made to the codebase without explicit human approval.

How it works

It acts as a local-first orchestration layer for existing AI coding agent CLIs (such as Claude Code, Codex, and Gemini CLI). The app captures a "frozen" snapshot of a task and distributes it to selected agents. It then structures the responses as evidence trails (Finding $\rightarrow$ Evidence $\rightarrow$ Decision $\rightarrow$ Disposition) rather than raw terminal output, gating any actual file writes behind a human adoption gate.

Who it’s for

Developers who use multiple AI coding agents and want to compare different implementation approaches, conduct independent reviews of code or research, and maintain a traceable audit trail of agent decisions before applying changes to their workspace.

Highlights

  • Multi-Agent Coordination: Supports 12+ local agent CLIs including Claude Code, Gemini CLI, and Qwen Code.
  • Collaboration Modes: Offers Direct sessions, Concurrent Responses for side-by-side comparison, and Auto Discussion V4 for multi-round negotiation and verification.
  • Human-in-the-Loop: Prevents automatic workspace writes, requiring user approval to adopt results.
  • Evidence-Based Review: Transforms raw CLI output into structured evidence and decision records for better inspectability.

Related

  • Project
    Aisland-SJL/dsh-worktable

    An agent-project workbench for DeepSeek Harness that provides a dockable multi-window workspace and a real-time control room for monitoring AI projects.

  • Project
    EleutherAI/sparsify

    A lean library for training and enable loading k-sparse autoencoders (SAEs) and transcoders on HuggingFace language model activations to aid in mechanistic interpretability.

  • Project
    franzenzenhofer/big-arrow-on-the-screen

    A macOS command-line tool that lets AI agents draw visible arrows and labels on the screen to guide users to specific UI elements or coordinates.

  • Project
    Qiuner/birdview

    A skill for AI coding agents that visualizes project architecture and constraints, requiring user confirmation of the change plan before the agent implements code edits.