Malt — Testing skill for Claude Code

MALT — Multi-Agent Local Team

**A local coordination service for independent coding agents, with durable inboxes and an orchestrator-controlled work/wait loop.**

MALT lets Claude Code, Codex, and other MCP-capable sessions communicate as members of one team. Each agent keeps its own model, context and terminal. A shared broker remembers their identities, messages, work requests and idle decisions in each target project's .malt folder. One broker serves a separate team for each project folder, so sessions in one repository never see the lead, workers or messages of another.

You explain the objective to the lead. Workers ask that lead for work, and the lead either assigns instructions or explicitly tells a worker it is not needed right now. Workers can wait for a response without repeatedly asking the model to inspect an empty inbox.

This repository includes durable communication, **cooperative wait/resume**, project-local storage, GitHub coordination, a managed factory supervisor and an installable PWA. A live subscription-authenticated Codex/Claude pilot delivered a narrowly scoped PR through two QA reviews, final approval, an expected-head merge and actual merged-main verification. This explicitly authorized private-repository pilot used local gates; live GitHub protected-review enforcement remains unverified. See [Mission control and human skill governance](docs/operator-console.md) for startup, model routing, local overrides and current limits. Attached MCP sessions still use the cooperative loop; managed factory sessions are launched and retired by their supervisor.

See [Project coordination and learning](docs/software-factory.md) for project-scoped GitHub setup, atomic parent/sub-issues, the learning steward, shared Claude/Codex skills, migration and the remaining factory milestones.

The public product and native command are **MALT** (`malt`). Existing `agentmesh` registrations and MCP tool names remain compatible. See [the branding and storage migratio