AI Team OS – a self‑driving AI‑company platform

What it is – AI Team OS (sometimes called AI Team OS or AI Company) is an open‑source, full‑stack system that turns a Claude‑style LLM (the “CEO”) into an autonomous, continuously‑running “company”. You, the human, act only as the Chairman: you set high‑level vision and approve occasional blockers. The platform then orchestrates dozens of specialist agents, a task wall, meetings, research pipelines, and a memory layer so work proceeds without you having to keep prompting the model.

---

Core ideas

Feature	What it does (in plain language)
Cross‑session orchestration	A single Claude session can watch and drive its sibling sessions, letting agents resume work, reuse existing sessions, or spawn new ones.
Memory System v2	Two‑layer memory: a direction layer injects your preferences and corrections into every new agent at birth; an episodic layer stores task‑level memos searchable with a fast BM25 index (Chinese‑bigram aware).
Progressive tool‑loading governance	The system keeps the number of loaded tools small by automatically rotating the most‑used ones (max 5) and offering read‑only or domain‑specific toolsets.
Workflow observability	Every Claude Workflow run is automatically turned into a “team” that is visualised on a live dashboard with swim‑lane timelines, token usage, and stall detection.
Ecosystem research platform	A curated knowledge‑base that ingests GitHub repos, summarises them, runs architecture analysis, debates findings, and can turn a research outcome into a concrete implementation task.
Knowledge layer	All memos, reports, and tasks are indexed in a reference graph (pure regex extraction, zero‑LLM cost) and a unified search that blends full‑text BM25, graph fan‑out, and exact‑ID matches.
Autonomous CEO loop	The CEO continuously pulls the highest‑priority item from the task wall, assigns the right specialist agent, parks blocked work, and launches R&D cycles when idle.
File‑truth source of truth	Model availability, leader liveness, and workflow telemetry are derived from on‑disk files and transcript timestamps rather than trusting agents’ self‑reports.
Model governance	Auto‑discovers every model that has ever appeared in your Claude transcripts, lets you set a global default model safely, and exposes the info via REST and MCP endpoints.
Team collaboration	25 ready‑made agent templates (engineering, testing, research, management, debate roles, etc.), channel‑style communication (team:, project:), unread badges, and a structured debate mode.
Safety & guardrails	Built‑in pattern detection, PII warnings, and mandatory team‑name declarations for non‑readonly agents to prevent rogue behaviour.

---

How you would use it

1. Install – the repo ships a FastAPI backend (Python 3.11+) and a React 19 dashboard. It also provides a set of MCP (Model Context Protocol) tools for programmatic access.
2. Start the CEO – launch the FastAPI server; the CEO session begins automatically, reading the ~/.claude/projects/ files to discover the current leader model.
3. Set vision – via the web UI or an MCP call you add high‑level goals to the task wall.
4. Let it run – the CEO continuously picks the next task, spawns the appropriate agent template, and updates the dashboard. If an agent hits a blocker that needs your approval, a message appears on the task wall; you approve or reject, and the CEO moves on.
5. Inspect – the Decision Cockpit shows every decision trace, the Workflow page shows live run timelines, and the Search bar lets you retrieve any past memo, report, or code change.
6. Extend – add custom MCP tools, new agent templates, or plug in additional LLM back‑ends; the system’s tool‑loading governance will automatically rotate the most‑used ones.

---

Who it’s for

• Technical founders or CTOs who want an AI‑first “company” that can keep working on product road‑maps without daily prompting.
• Research teams that need a reproducible, auditable pipeline for scanning external codebases, debating design choices, and turning findings into implementation tasks.
• Developers who like to experiment with multi‑agent orchestration, memory persistence, and zero‑LLM knowledge graphs.

---

Limitations (as described in the README)

• The platform assumes you are using Claude‑style sessions (claude -p …). Other LLM back‑ends would need adapters.
• Some advanced features (e.g., the prompt‑time badge for Codex) are still under measurement and not fully verified.
• The system relies heavily on local transcript files; if those are missing or corrupted, model discovery and leader probing may fail.

---

Bottom line – AI Team OS is a comprehensive, open‑source “operating system” for running a self‑driving AI company. It combines multi‑agent orchestration, persistent memory, research pipelines, and a rich observability dashboard, letting a human set direction while the AI continuously executes, learns, and improves.