Loom – Delivery Harness for Coding‑Agents

What it is – Loom is a local runtime that sits next to a code‑generation agent (Claude Code, Codex, OpenCode, …). It turns a single LLM‑generated suggestion into a full delivery loop: clarify requirements, plan, write code, run tests, review, repair failures, preview the result and hand‑off the work. All state is kept in a hidden .loom/ folder so a session can be resumed without re‑reading the whole repo.

Who it helps – Developers who already use an AI coding assistant and need more reliable, repeatable deliveries – e.g. building a new feature, reviewing a PR, or deploying a preview. Loom adds the missing orchestration and bookkeeping that a raw LLM prompt lacks.

How it works

1. Plugin registration – An installer registers a small MCP (Message‑Control‑Protocol) plugin with the chosen agent (@loom for Codex, /loom for Claude Code/OpenCode).
2. MCP server – A Rust‑based MCP runtime runs locally, exposing a state‑machine that knows the delivery stages.
3. Delivery steps – When you type @loom build … the agent asks Loom for a plan, Loom returns a manifest of read/write groups and verification goals, the agent generates code, Loom validates the result, records evidence, and either moves to the next stage or asks for a repair.
4. Context routing – Only the minimal required repository slices, test results, and knowledge‑source snippets are sent to the LLM each turn, keeping prompts short and focused.
5. Knowledge sources (optional) – You can register local docs (MD, PDF, JSON, etc.) that Loom indexes and queries during requirement clarification.
6. Resume & hand‑off – All artifacts, task contracts, review notes and repair requests are persisted under .loom/. Running @loom continue resumes exactly where you left off, even after closing the editor.

Key features

• Multi‑agent support – Works with Codex, Claude Code, OpenCode, and future coding agents.
• Rust MCP core – Fast, small runtime (migrated from TypeScript in July 2026).
• Docker‑based preview – @loom deploy builds a local Docker‑Compose preview with logs and runtime facts.
• Recursive self‑improvement – Test/repair feedback loops automatically feed into the next iteration.
• Extensible knowledge bases – Add product rules, design standards, etc., that the agent can query during clarification.
• Open‑source Apache‑2.0 – Free to modify and integrate.

Getting started

# Install for all supported agents (Linux/macOS)
curl -fsSL https://github.com/valkor-ai/loom/releases/latest/download/install.sh \
  | bash -s -- --agent all

Then, inside a project opened with your coding agent:

@loom build "add a team settings page"   # Codex
/loom build "add a team settings page"   # Claude Code / OpenCode

Use @loom continue, @loom review, and @loom deploy to move through the pipeline. The installer also provides Windows PowerShell scripts and a --local-build mode for developers who want to compile the Rust server themselves.

When to use it – When a single LLM output isn’t enough: complex features, multi‑step refactors, or any work that must survive interruptions and be handed off to another developer or another LLM session.

---

Loom is a genuine software project focused on AI‑assisted software delivery, fitting squarely within the frontier of coding‑agent tooling.