valkor-ai/loom

What it solves

Coding agents often struggle with large software tasks because they lose context, skip critical checks, or fail to resume work after a session ends. Loom provides a stateful delivery harness that prevents goal drift and partial completions by turning one-shot coding sessions into a structured, resumable loop of planning, building, testing, and repairing.

How it works

Loom operates as a local Model Context Protocol (MCP) delivery state machine that runs alongside agents like Claude Code, Codex, or OpenCode. Instead of the agent managing the entire workflow from memory, Loom manages the project-local state in a .loom/ directory. It routes the agent through a specific sequence: clarifying requirements (optionally using local knowledge sources), establishing a technical baseline, executing bounded tasks with specific read/write boundaries, and validating results through a review and repair loop. It uses "context routing" to provide the agent with only the necessary field groups and references for the current step, reducing noise and token usage.

Who it’s for

Developers using AI coding agents who need to move beyond simple prototypes to shipping verifiable software, particularly for tasks too large for a single prompt such as feature development, deployment, or complex repairs.

Highlights

• Stateful Delivery: Saves requirements, task progress, and repair notes to allow sessions to be resumed via a continue command.
• Context Routing: Prevents context window overflow by delivering targeted retrieval paths instead of reloading entire repositories.
• Knowledge Integration: Allows users to register local domain documents as searchable knowledge sources to aid in requirement clarification.
• Multi-Agent Support: Works across different MCP-capable agents including Claude Code, Codex, and OpenCode.
• Local Deployment Previews: Prepares Docker Compose previews with topology-aware services and health checks for validation.