Wrap a coding agent in a structured delivery loop for features bigger than a one-shot edit.
Resume multi-session agent work from saved project state instead of starting over after an interruption.
Add separate verification and repair steps so an agent's work is checked instead of self-graded.
Produce a human-readable delivery report and preview evidence when handing off finished work.
| valkor-ai/loom | sarthakrawat-1/sketchbook-ui | punkpeye/mcp-proxy | |
|---|---|---|---|
| Stars | 261 | 265 | 267 |
| Language | TypeScript | TypeScript | TypeScript |
| Last pushed | — | — | 2026-06-27 |
| Maintenance | — | — | Active |
| Setup difficulty | — | easy | easy |
| Complexity | — | 2/5 | 3/5 |
| Audience | developer | developer | developer |
Figures from each repo's GitHub metadata at analysis time.
Loom is an open-source delivery framework that wraps around coding agents like Claude Code, Codex, and OpenCode to make them behave like repeatable software delivery systems rather than one-shot prompt tools. It does not replace the agent or editor a developer already uses. Instead, it turns each delivery goal into a structured loop covering planning, building, verification, repair, preview, and final handoff. The core problem Loom addresses is that long-running agent work tends to fail in predictable ways. An agent might declare a task done after only partial progress, drift away from the original goal over many turns, judge its own work with bias since it wrote the code being reviewed, waste tokens re-reading the whole codebase repeatedly, or leave a confusing state for a human or another agent to pick up after an interruption. Loom responds to each of these with a specific mechanism. Bounded tasks and explicit result files stop agents from claiming completion too early. Saved scope, architecture contracts, and task plans keep the original objective intact across sessions. Separate review, verification, and repair steps keep an agent's self-check honest by treating implementation and validation as distinct stages. Project summaries and stored test results cut down on repeated full-repository reads. Delivery reports, preview checks, and repair history leave behind a record that is readable by both humans and other agents. Rather than being a set of static prompt files or a single fixed workflow, Loom chooses a delivery path dynamically for each goal, then persists project context, task state, backend state, test results, and handoff reports in a .loom/ folder so work can resume from the right step even after a session ends or the underlying agent adapter changes. It exposes an agent-neutral command line interface, so it can sit in front of different coding agents rather than being tied to one. Loom is intended for requests bigger than a quick one-off edit, such as a feature that needs clarification, architecture planning, task splitting, review, repair, preview, deployment, and a clean handoff. It is built with TypeScript on Node.js 20 or newer and released under the Apache 2.0 license, and its README is also available in Simplified Chinese.
Loom is an open-source delivery framework that wraps coding agents like Claude Code and Codex in a persistent plan, build, verify, repair, and handoff loop so work survives interruptions.
Mainly TypeScript. The stack also includes TypeScript, Node.js, Claude Code.
Apache 2.0 license: use freely for any purpose, including commercial use, as long as you keep the copyright and license notices.
Mainly developer.
This repo across BitVibe Labs
Verify against the repo before relying on details.