A delivery state machine for coding agents that remembers where it left off

What it does

Loom is an open delivery harness that runs as a local MCP server (Rust) alongside your existing coding agent. Instead of letting the agent freestyle a large change, Loom drives a structured loop: clarify scope, plan, build, run checks, repair failures, preview, and leave a handoff. All state — requirements, task progress, results, repair history — lives in a project-local .loom/ directory, so a fresh session picks up the actual state of the work rather than reconstructing it from chat scrollback.

The interesting bit

The agent doesn’t decide the workflow from memory; it asks Loom for the next request, reads only the declared fields, writes the expected artifact, and submits it back. That’s a state machine wrapped around an LLM, which is a refreshingly honest way to handle agents that forget what they were doing two turns ago. Context routing also selects task-specific references instead of dumping the whole repo into every turn.

Key highlights

• Works with Claude Code, Codex, and OpenCode through a unified @loom / /loom command surface (build, continue, review, deploy, status).
• continue is the universal “I don’t know what step is next” action — safe after interruptions or reopened sessions.
• Optional knowledge sources: index local docs (.md, .pdf, .docx, and more) so requirement clarification can pull in product rules and design standards.
• MCP runtime rewritten from TypeScript to Rust for a smaller, faster core; algorithms run in an isolated Python runtime.
• Installer verifies package checksums and refuses to overwrite files it can’t prove it owns — a rare courtesy.

Caveats

• loom deploy requires Docker; the rest runs locally.
• The README is heavy on command surface and light on internals — the “how it works” section cuts off mid-sentence, so you’ll need the linked technical report for the full protocol.
• Knowledge builds can be interrupted and need an explicit resume — semantic indexing is apparently not a one-shot affair.

Verdict

Worth a look if you use coding agents for multi-session feature work and are tired of re-explaining the plan every time you reopen a chat. If your agent usage is one-shot scripts and small edits, this is machinery you don’t need.