CronusL-1141/AI-company

What it solves

AI Team OS solves the problem that AI coding agents like Claude Code and Codex lose all context when a session ends. When you close a terminal or finish a chat, the tasks, memos, decisions, and history disappear along with it. This project gives those agents a shared, persistent place to record work so any future session can pick up where the last one left off, even if it uses a different agent tool.

How it works

AI Team OS runs locally as a layer on top of your chosen agent host. You still use Claude Code or Codex to do the actual work, but the OS provides MCP tools, hooks, and a web dashboard to keep records. When a session starts, it resolves the project, reads the task wall and memory, and a session Leader coordinates authorized work. As agents run, they log progress, memos, and reports through shared tools. Other sessions pull that context back in. The OS also includes a memory system with a direction layer for preferences and an episodic layer for task evidence, plus a knowledge graph built from references in your notes. Everything is stored locally and searchable without needing extra model calls for retrieval.

Who it’s for

This is built for developers and teams who work with Claude Code, Codex, or both, and need to keep multiple parallel or sequential agent sessions coordinated. It’s especially useful for people who hand off work between sessions, run long projects with lots of context, or want transparency and audit trails around what their agents did.

Highlights

• Escape hatches: lets Claude Code and Codex share the same task wall, memory, reports, and channels while keeping their native tools and permissions separate.
• A task wall with dependencies, blockers, and completion records, plus a live dashboard with team activity.
• A memory system with a direction layer (preferences, corrections) and an episodic layer (task memos), all retrievable via BM25 search without extra LLM calls.
• Progressive tool-loading governance: dynamically rotates a hot-tool whitelist based on real usage frequency, and supports read-only and toolset-group profiles.
• Automatic tracking of Claude Code Workflow runs with live telemetry and calibrated stall detection.
• An ecosystem research platform that accumulates knowledge about open-source repos through staged summaries, architecture analysis, debates, and integration marking.
• A knowledge layer with a reference graph and unified search across task memos, reports, and tasks.
• 25 professional role templates and structured meeting templates for coordination.
• Full transparency: an event stream, decision cockpit, and what-if analyzer.
• Safety features like guardrails, watchdog patrols, and concurrency warning hooks.