English | 中文

AI Team OS

Your AI coding tool stops when you stop prompting. Ours doesn't.

🤝 Now on Codex as well. AI Team OS started life as a Claude Code plugin. A Codex session now plugs into the same operating system: the same task wall, memory, reports and Dashboard, the same MCP tools, with Codex's own native hooks feeding the observation layer. Two harnesses can work one project as one team and leave each other messages; when one names the other, the recipient sees it at the top of its next turn (measured on both the Codex CLI and the desktop app). The runtime fixes now on master were produced exactly this way: Codex wrote the branches, Claude Code reviewed and merged them, and the two coordinated over the OS channel throughout. Still Claude Code only for now: one-command install, the session-start briefing and the direction-layer memory handed to new subagents, and waking an idle session when mail arrives. Codex setup is manual today: register the adapter's hook scripts in Codex and point it at the same MCP server.

⚡ v1.12.3 - Codex now actually sees the unread line: its host does not inject a hook's plain stdout, only the structured additionalContext form, which is why a hook that ran and produced output had never been observed reaching the model. Measured after merge on both the CLI and the desktop app. Alongside it, the runtime work that had lived only in five uncommitted desktop worktrees is on master (slow-client isolation for WebSocket broadcast, an on-demand PID ledger that adopts only a verified process, one SQLite admission slot reserved for hook events, coalesced Dashboard refreshes), reviewed one dimension at a time with a refuter per finding. The one blocking finding: psutil was load-bearing and undeclared, passing every test only because an unrelated package brought it along; it is now declared, and its absence degrades instead of paralysing. And the README finally says the OS runs on Codex, with the boundary of what is still Claude Code only.

Full version history: CHANGELOG.md

Python License FastAPI React MCP Stars

116 MCP tools · 211 REST endpoints · 23 dashboard pages · 2,576 tests · 25 agent templates · 42 ecosystem research tools · 21 machine-checked invariants

AI Team OS turns Claude Code, and now Codex, into a self-driving AI company. You're the Chairman. AI is the CEO. Set the vision — the system executes, learns, and evolves autonomously.

The Problem With Every Other AI Tool

Every AI coding assistant works the same way: you prompt, it responds, it stops. The moment you step away, work stops. You come back to a blank prompt.

AI Team OS works differently.

You walk away at night. The next morning you open your laptop and find:

  • The CEO checked the task wall, picked up the next highest-priority item, and shipped it
  • When it hit a blocker that needed your approval, it parked that thread and switched to a parallel workstream
  • R&D agents scanned three competitor frameworks and found a technique worth adopting
  • A brainstorming meeting was organized, 5 agents debated 4 proposals, and the best one was put on the task wall

You didn't prompt any of that. The system just ran.

How It Works

You're the Chairman. The AI Leader is the CEO.

The CEO doesn't wait for instructions. It checks the task wall, picks the highest-priority item, assigns the right specialist Agent, and drives execution. When blocked, it switches workstreams. When all planned work is done, R&D agents activate — scanning for new technologies, organizing brainstorming meetings, and feeding improvements back into the system.

Every interaction makes the system understand you better. Memory System v2 distills your preferences and corrections into a team direction layer that every dispatched Agent inherits at birth — you never say the same thing twice, and no next Agent repeats a pit an earlier one already fell into.

Core Capabilities

1. Cross-Session Orchestration (new in v1.10.0)

A single CC session can now observe and drive its sibling sessions for one operational turn, instead of only being able to spawn brand-new ones:

  • Wake system v2: the /api/wake/actionable single-source predicate feeds both the event watcher and the turn-end guard; SessionStart moves from a fixed 30-minute cron to dynamic /loop intervals; a Stop-hook turn-end guard always lets decision:block and user-stop keywords pass through; a session-scoped event watcher carries a 1-hour hard timeout. No resident daemons.
  • Fleet downlink primitive: headless claude -p --resume <session_id> drives a target sibling session for one turn, reusing the existing wake machinery (semaphore, fuse, allowlist, per-session dedupe, full audit trail).
  • agent_reuse_recommend MCP tool: a three-way reuse decision (reuse / slim-then-reuse / spawn-new) scored by domain match, reachability (live / resumable / cross-session / expired), and context watermark.
  • Context watermark ledger: exact token usage read from the transcript tail (cheap-checks-first), surfaced as a three-color watermark bar on agent views and on the new fleet / worktree observability cards.
  • Compaction checkpoint (v1.11.0): PreCompact freezes the OS-side operating picture — agents in flight, open tasks, decisions queued for you — and SessionStart(source=compact) hands it straight back. CC's own summary body is deliberately not stored: after compaction it is already in the model's context; what a compacted Leader loses is the OS-side state it no longer knows to ask about.
  • CC session registry as a second liveness track (v1.11.0): ~/.claude/sessions/<pid>.json carries a real pid and CC's own idle/busy state, which distinguishes "process gone" from "process alive but quiet" — a distinction transcript freshness cannot make. It runs alongside the existing verdict and only records where the two disagree; the verdict itself is unchanged until the divergence data says otherwise.
  • Background daemon sessions are visible (v1.11.0): GET /api/hooks/background-jobs reads CC's own job state, so a --bg session that outlives its foreground window no longer looks like "nobody is working".

Usage guidance:

  • A new session's SessionStart briefing already points you at running /loop once - follow it instead of guessing at intervals.
  • Check the project detail page for the fleet card (per-session CEO / model / in-flight tasks / watermark) and the worktree card (branch ownership + unlanded-work status) before you act.
  • Call agent_reuse_recommend before dispatching a follow-up agent - reusing a live or resumable sibling session beats spawning a fresh one.
  • The S4 worktree teardown guard and per-template isolation: worktree defaults apply automatically; no configuration is needed.

2. Memory System v2 — two-layer memory, every Agent inherits at birth (new in v1.9.0)

The OS's signature differentiator: your team's preferences, corrections, and hard-won lessons flow automatically to every Agent it dispatches.

  • Direction layer (user preferences / corrections / design intent, 4 kinds): resident injection via both the SessionStart and SubagentStart hooks — every sub-Agent inherits the team's values and red lines the moment it's born, so you don't repeat yourself. The size guardrail is a single axis: storage cap = injection budget (per-bucket character quotas, global 1200 + 1500 per project + user 300 = 3000 chars, <=400 chars per entry), so whatever fits is what actually ships; a full bucket hands back its complete contents and demands a cleanup before the retry. supersedes swap to prevent bloat, invalidate-never-delete for auditability. Writes are scanned for invisible Unicode, instruction-override phrasing, and credential shapes — the direction layer lands in every Agent's system prompt, which makes it an injection amplifier.
  • Episodic layer (task_memos ledger): task-level execution memos promoted to a dedicated table (row IDs / invalidation axis / quality score / scope_path), recalled on demand via pure-Python BM25 Chinese retrieval; 123 legacy memos backfilled with zero loss.
  • On-demand reconcile (memory_reconcile): zero-LLM BM25 candidate clustering, then merge / invalidate / score / distill on agent confirmation — "the agent computes, the tool persists", with no background resident process introduced.

Surfaces: MCP memory_add / memory_list / memory_invalidate / memory_search / memory_reconcile_candidates / memory_reconcile_apply.

3. Progressive Tool-Loading Governance (new in v1.9.0)

Treats the resident context budget as the scarce resource it is — however many tools exist, they never drown your Agent.

  • alwaysLoad dynamic rotation: at session start a single SQL recomputes the hot-tool whitelist by 7-day real call frequency (>=2-day span gate against bursty spikes + 20% hysteresis, hard cap <=5), and CC skips ToolSearch for them. Not additive, not hand-tuned; any stats failure silently degrades to all-defer, and every whitelist is logged for audit.
  • AITEAM_TOOLSETS group switch: 16 capability-domain toolsets; a startup env var decides which modules register. default core profile = task/team/memory/infra/reports (29 tools, hard cap <=50), with incremental default,ecosystem — fits non-CC clients that cap tool counts.
  • AITEAM_READONLY read-only profile: an orthogonal overlay that strips every write tool by explicit allowlist and keeps only read tools — ideal for audit / observer sessions.
  • 5 templates on least privilege: meeting-facilitator / debate advocate & critic / technical-writer / project-manager carry disallowedTools structural denials; engineering / testing templates untouched.

4. Workflow / ultracode Persistent Observability (v1.7.0)

The OS does not intercept CC's built-in ultracode/Workflow — it becomes its persistent governance layer. Every Workflow run is automatically tracked into the OS, with no manual team setup:

  • Auto-tracking: a hook turns each Workflow run into an OS "team" (workflow-<wf_id>) the moment it starts
  • Dashboard /workflows: a live feed of run cards, a phase swimlane timeline, and per-agent telemetry — tokens / duration / status / tool-call counts, advancing live via incremental journal tailing while a run executes
  • Calibrated stall detection: the stall threshold was calibrated on 3,378 real agent intervals (p99 = 77.6s, longest healthy silence 173.8s) and set at 5.2× the worst healthy case — it flags late rather than crying wolf
  • Project-detail integration: workflow team rows carry an inline run summary (status / agent count / duration / finish time) plus a "view swimlane" deep link; members display semantic phase labels (e.g. audit:sourceA) instead of ids
  • Leader auto-detection: a project's Leader session / model / liveness is probed directly from the ~/.claude/projects/ file truth by the backend — zero registration dependency, /model switches surface in real time
  • MCP tools: workflow_list (browse runs), workflow_get (full archive + per-agent rows), workflow_reconcile (repair from on-disk snapshots after the OS was offline)
  • Self-healing ingestion: hook receipt anchors + on-disk snapshot reconciliation + a reaper backstop close offline gaps automatically — finished runs on disk are ingested idempotently; cross-project attribution matches the on-disk path slug against registered projects

5. Ecosystem Research Platform — 42 tools

A project-isolated knowledge base that accumulates research findings over time. Each repo progresses through 4 stages (a progressive funnel, since v1.5.0), with token-efficient triggers and append-only history:

  • Stage 0 — Auto shallow-summary on archive: newly-archived repos automatically get a 200-400 char ai-engineer summary (core function / positioning / advantages). 8-class failure handling with self-learning hooks (3+ same-class fails surface through self_learning_pending; the queue exposes recorder/searcher injection points you can wire to your own lesson store)
  • Stage 1 — On-demand architecture analysis: user picks research direction ("memory_system") → batch-dispatch backend-architect agents to read architecture key files
  • Stage 2 — Multi-perspective debate: triggers existing debate_start (NOT a built-in debate engine — reuses meeting system)
  • Stage 3 — Reference / Integrate marking: mark_as_reference adds tag for future quick recall; start_integration triggers existing task_create for actual implementation
  • Active vs Full dual-view: data is append-only forever. Stars-falling repos kept (just is_active=False); stars climbing back auto-promotes + re-queues Stage 0
  • Dashboard /ecosystem: list with stage badges + research timeline + project filter dropdown + candidate-filter page (/ecosystem/research) + per-project settings tab — the single largest tool family in the OS

6. Knowledge Layer — Reference Graph + Unified Search (v1.8.0)

Everything the OS records — task memos, reports, tasks — becomes recallable knowledge:

  • Reference graph (P1a): a zero-LLM regex extractor mines OS-native ID references (wf_id / commit hash / task uuid / [[memory]]) out of memos and reports into an append-only knowledge_links table — the graph is a derived view, rebuildable from source text at any time
  • Unified search (P1b): /api/search fuses three arms via RRF — BM25 full-text (Chinese bigram native), knowledge-graph fanout (an ID query pulls in everything linked to it), and exact ID-prefix / title match
  • Global search box in the Dashboard header, plus MCP tools unified_search / link_query / link_trace — recall past work by natural language ("how was the attribution fix done"), a wf_ id, or a commit hash

Why zero-LLM? The graph is a derived view: plain regexes extract the IDs, the whole graph can be rebuilt from source text at any time, and both extraction and retrieval cost zero tokens. Your recall pipeline never touches your model budget.

7. Task Wall · Meetings · 22-Page Dashboard

Governance ledger and panoramic visualization — everything leaves a trace:

  • Task wall: a live board of pending / in-progress / done, event-driven + intelligent Agent matching + deadlock detection
  • 8 structured meeting templates (keyword auto-select, built on Six Thinking Hats / DACI / Design Sprint) — every meeting must produce an actionable conclusion; "we discussed but didn't decide" is not an outcome
  • 22-page React 19 Dashboard: Command Center / /workflows swimlane / decision timeline / meeting room / Ecosystem suite / Model Governance Settings

8. Autonomous Operation

The CEO never idles. It continuously advances work based on task wall priorities:

  • Checks the task wall for the next highest-priority item when a task completes
  • When blocked on something requiring your approval, parks that thread and switches to parallel workstreams
  • Batches all strategic questions and reports them when you return — no interruptions for tactical decisions
  • Deadlock detection: if the loop stalls, it surfaces the blocker rather than spinning

And it doesn't just execute — it evolves:

  • R&D cycle: research agents scan competitors, new frameworks, and community tools; findings go to brainstorming meetings where agents challenge each other; conclusions become implementation plans on the task wall

9. File Truth as Source of Truth

Most multi-agent stacks trust agents to register themselves and self-report their status. AI Team OS treats self-reports as claims and files as facts — three subsystems already run on this philosophy:

  • Leader probing: a project's Leader session, model, and liveness are read straight from ~/.claude/projects/ — transcript mtime is liveness, the model name in the transcript tail is the model. We don't ask an agent which model it runs — what's read out of the transcript is what's true.
  • Model discovery: "available models" = every model that has actually appeared in your CC transcripts. Zero API dependency, zero hardcoded list — a hardcoded list will never contain your third-party gateway model; a transcript scan can't miss it.
  • Workflow telemetry: on-disk run files are the full telemetry truth; the OS's projection tables are rebuildable caches of immutable files. Attribution iron law: a run belongs to a project only when its on-disk path slug exactly matches the registered project root — never guessed.

10. Model Governance (v1.8.1)

Know which models you can actually launch — and control what your sessions start on:

  • Auto-discovery of genuinely available models: scans every CC transcript on your machine in about a second (60s cache) — including third-party gateway models that no hardcoded list would ever ship
  • One-click global default startup model: written to ~/.claude/settings.json under triple write protection — touches only the model key, keeps a .bak-aiteam backup, writes atomically, refuses corrupted files
  • Zero coercion: soft reminders only, never a block — and CC Workflow runs are fully exempt

Surfaces: REST /api/models/{available,default} · MCP model_config_get / model_config_set · the Model Governance card in Dashboard Settings.

11. Team Collaboration

Not a single Agent. A structured organization:

  • 25 professional Agent templates (23 base + 2 debate roles) with recommendation engine — Engineering, Testing, Research, Management — ready out of the box
  • Department grouping — Engineering / QA / Research with cross-team coordination
  • Channel communication: team: / project: / global channels with @mention support
  • Unread badge across sessions (new in v1.12.0): two AI sessions can leave each other messages, and the recipient finds out. On CC, measured end to end: a UserPromptSubmit hook injects one line when someone has named you, channel_read_ack clears it, and the next turn recomputes to zero so the line disappears on its own; arm the session-scoped watcher with a reader identity (bash scripts/os-watch.sh <sid> <team> <reader> &) and a new message wakes the session in about 8 seconds with no user input at all. Sending, reading and acknowledging also work from Codex, and its prompt-time line is now measured too (2026-09-09, on both the CLI and the desktop app): the hint reaches the model's context before any tool runs, once the hook emits it as hookSpecificOutput.additionalContext rather than plain stdout, which Codex's host does not inject. What differs is receiving without a prompt: CC has the message wake the session (event-driven, the watcher polls every 8 seconds), while Codex can hold a channel_wait open and return when a message lands, with delivery_source saying whether it arrived by push or by the final read; an idle Codex session with nothing waiting is not woken. Neither side can reach a session that has already exited: nothing is running there to receive. The os-channel skill carries the exact calls and a per-harness capability table separating what is measured from what is not.
  • Debate mode: 4-round structured debate (Advocate→Critic→Response→Judge) via debate_start / debate_code_review
  • Cross-agent lessons: failure_analysis writes root-cause antibodies into project memory, and every dispatched sub-Agent inherits them through the direction layer

12. Full Transparency

Nothing is a black box:

  • Decision Cockpit: event stream + decision timeline + intent inspection — every decision has a traceable record
  • Activity Tracking: real-time status of every Agent and what it's working on
  • What-If Analyzer: compare multiple approaches before committing, with path simulation and recommendations

13. Safety & Behavioral Enforcement

Built-in guardrails so the system can run unsupervised without surprises:

  • Guardrails L1: 7 dangerous pattern detections + PII warnings + InputGuardrailMiddleware
  • Local agent blocking: all non-readonly agents must declare team_name/name — prevents rogue background agents
  • S1 safety rules: regex-based scan catches destructive commands (rm -rf, force push, hardcoded secrets) including uppercase flags and heredoc patterns
  • 4-layer defense rule system: 48+ rules covering workflow, delegation, session, and safety layers
  • Concurrent-edit warnings: hooks flag a file two agents touched in quick succession, read straight from recent edit events (the cooperative file-lock tools were retired in v1.10.3 — the lock file was empty in every real run)
  • Agent Watchdog: on-demand POST /api/teams/{id}/watchdog/check plus the background patrol — flags BUSY-timeout agents, long-pending tasks and unblockable dependencies
  • Self-patrol: watchdog lease patrol + reaper reconciliation backstop + identity verification before any kill — the OS keeps eyes on itself, not just on your agents
  • Completion verification: verify_completion checks task status + memo existence — prevents hallucinated "done" reports
  • Ecosystem integration recipes: 4 preset recipes (GitHub / Slack / Linear / Full-stack team) under find_skill(level=2, category="integration")
  • find_skill 3-layer progressive discovery: quick recommend → category browse → full detail, reducing tool-call overhead

14. Zero Extra Cost

Runs entirely within your existing Claude Code subscription:

  • No external API calls, no extra token spend
  • MCP tools, hooks, and Agent templates are all local
  • The memory system and knowledge layer are zero-LLM by design — direction-layer injection, graph extraction, search, and reconcile coarse-pass all cost zero tokens
  • 100% utilization of your CC plan

More Capabilities (legacy & secondary — still running, queryable on demand)

  • Failure Alchemy: failure_analysis still runs as part of the loop subsystem — every failed task extracts root cause and produces Antibody (stored in team memory to prevent repeats) / Vaccine (high-frequency failures become pre-task warnings) / Catalyst (analysis injected into future Agent system prompts). No longer the headline, but defensive rules keep accruing.
  • AWARE loop memory · find_skill 3-layer discovery (skills + integration recipes) · Prompt Registry: see the full tool table below. The scheduler and the loop state machine were retired in favour of CC-native Cron* and on-demand tools (CC-is-not-always-on principle); the wake_agent schedule kind survives for the fleet wake subsystem.

It Built Itself

AI Team OS manages its own development — and since v1.7.0, it can prove it with its own telemetry:

  • Every feature line from v1.7.0 to v1.9.0 — the observability layer, the knowledge layer, model governance, Memory System v2, tool-loading governance — shipped through CC Workflow runs that the OS tracked itself. Open /workflows and replay how the system built its own features, swimlane by swimlane.
  • Competitive research across CrewAI, AutoGen, LangGraph, and Devin feeds the roadmap through multi-agent brainstorming meetings — the minutes live in the OS's own report store.
  • It learns from its own incidents, too: every machine-checked invariant in scripts/check_invariants.sh was distilled from a real accident in this repo's history.

The system that builds your projects... built itself. With receipts.

How It Compares

Dimension AI Team OS CrewAI AutoGen LangGraph Devin
Category CC Enhancement OS Standalone Framework Standalone Framework Workflow Engine Standalone AI Engineer
Integration MCP Protocol into CC Independent Python Independent Python Independent Python SaaS Product
Memory System Two-layer: direction layer inherited at birth + episodic BM25 ledger + on-demand reconcile Short-term context Short-term context Checkpoint state In-session
Tool-Loading Governance alwaysLoad rotation + group switch + read-only profile + template least-privilege None None None None
Autonomous Operation Continuous loop, never idles Task-by-task Task-by-task Workflow-driven Limited
Meeting System 8 structured templates with auto-select None Limited None None
Failure Learning Failure Alchemy (Antibody/Vaccine/Catalyst) None None None Limited
Decision Transparency Decision Cockpit + Timeline None Limited Limited Black box
Workflow Observability Swimlane timeline + per-agent telemetry + offline reconcile over CC Workflow None None Graph state only None
State Source File truth — transcripts / journals read directly Agent self-report Agent self-report In-process state Black box
Rule System 4-layer defense (48+ rules) + behavioral enforcement Limited Limited None Limited
Agent Templates 25 ready-to-use + recommendation engine Built-in roles Built-in roles None None
Dashboard React 19 visualization Commercial tier None None Yes
Open Source MIT Apache 2.0 MIT MIT No
Claude Code Native Yes, deep integration No No No No
Extra Cost $0 (CC subscription only) API costs API costs API costs $500+/mo

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                     User (Chairman)                              │
│                         │                                       │
│                         ▼                                       │
│                   Leader (CEO)                                   │
│            ┌────────────┼────────────┐                          │
│            ▼            ▼            ▼                          │
│       Agent Templates  Task Wall  Meeting System                 │
│      (25 roles)       auto-assign  (8 templates)                 │
│            │            │            │                          │
│            └────────────┼────────────┘                          │
│                         ▼                                       │
│              ┌──────────────────────┐                           │
│              │   OS Enhancement Layer│                           │
│              │  ┌──────────────┐    │                           │
│              │  │  MCP Server  │    │                           │
│              │  │ (113 tools)  │    │                           │
│              │  └──────┬───────┘    │                           │
│              │         │            │                           │
│              │  ┌──────▼───────┐    │                           │
│              │  │  FastAPI     │    │                           │
│              │  │  REST API    │    │                           │
│              │  └──────┬───────┘    │                           │
│              │         │            │                           │
│              │  ┌──────▼───────┐    │                           │
│              │  │  Dashboard   │    │                           │
│              │  │ (React 19)   │    │                           │
│              │  └──────────────┘    │                           │
│              └──────────────────────┘                           │
│                         │                                       │
│              ┌──────────▼──────────┐                            │
│              │  Storage (SQLite)   │                            │
│              │  + WAL journaling   │                            │
│              │  + Memory System    │                            │
│              └─────────────────────┘                            │
└─────────────────────────────────────────────────────────────────┘

Five-Layer Technical Architecture

Layer 5: Web Dashboard    — React 19 + TypeScript + Shadcn UI (23 pages)
Layer 4: CLI + REST API   — Typer + FastAPI
Layer 3: Team Orchestrator — LangGraph StateGraph (optional extra — CLI graph execution only)
Layer 2: Memory Manager   — SQLite-backed store + pure-Python BM25 retrieval
Layer 1: Storage          — SQLite (WAL journaling) · PostgreSQL support on the roadmap

Hook System (13 scripts across 15 Lifecycle Events — The Bridge Between CC and OS)

SessionStart     → auto_install.py, session_bootstrap.py, send_event.py
                   — Auto-install deps + inject Leader briefing / core rules / team state
SubagentStart    → inject_subagent_context.py, send_event.py   — Inject sub-Agent OS rules (2-Action etc.)
SubagentStop     → send_event.py                 — Record sub-Agent lifecycle event
PreToolUse       → workflow_reminder.py, send_event.py
                   — Workflow tracking reminders + event forwarding
PostToolUse      → workflow_reminder.py, deep_review_link.py,
                   meeting_ecosystem_writeback.py, send_event.py
TaskCompleted    → cc_task_bridge.py             — Mirror finished CC tasks onto the OS wall (owned or dependency-linked ones only)
TeammateIdle     → send_event.py                 — CC's own teammate-idle signal, recorded alongside the OS liveness track (observation only, changes no status)
UserPromptSubmit → context_tracker.py            — Track context usage
                 → channel_unread.py             — Unread channel badge
                 → turn_end_guard.py             — Standby reminder (user-prompt mode)
SessionEnd       → send_event.py                 — Record session end event
Stop             → send_event.py                 — Record stop event
PermissionDenied → permission_denied_recovery.py — Permission-denied self-recovery
PreCompact       → pre_compact_save.py           — Freeze the OS-side battle state (in-flight agents / open tasks / pending decisions) into a checkpoint
PostCompact      → send_event.py                 — Confirm the compaction actually happened (a triggered compaction can still be cancelled)
WorktreeRemove   → send_event.py                 — An isolated worktree is gone

Quick Install (AI-Assisted)

Tell Claude Code:

"Read https://github.com/CronusL-1141/AI-company/blob/master/INSTALL.md and follow the instructions to install AI Team OS"

Claude Code will read the install guide and walk you through the setup automatically.

Important: Install AI Team OS to your system Python, not inside a project virtual environment. If installed in a venv, AI Team OS will only work in that specific project. Run deactivate first if a venv is currently active, then install.

Quick Start

Prerequisites

  • Python >= 3.11
  • uv (pip install uv)
  • Claude Code (MCP support required)
  • Node.js >= 20 (Dashboard frontend, optional)

Option A: Plugin Install (Recommended — for most users)

# Install uv (Python package runner, required for MCP server)
pip install uv

# Add marketplace + install plugin
claude plugin marketplace add CronusL-1141/AI-company
claude plugin install ai-team-os

# Restart Claude Code — first launch takes ~30s to set up dependencies
# Subsequent launches are instant

# Update to latest version anytime
claude plugin update ai-team-os@ai-team-os

Note: First launch after install takes ~30 seconds while dependencies are automatically configured. This only happens once — subsequent sessions start instantly with 116 MCP tools ready.

Option B: Source Install (for developers — editable, tracks latest source)

# Step 1: Clone the repository
git clone https://github.com/CronusL-1141/AI-company.git
cd AI-company

# Step 2: Run the installer (auto-configures MCP + Hooks + Agent templates + API)
python3 install.py

# Step 3: Restart Claude Code — everything activates automatically
# API server starts automatically when MCP loads. No manual startup needed.
# Verify: run /mcp in CC and check that ai-team-os tools are mounted

Dependencies: greenlet (needed by SQLAlchemy async on Apple Silicon) is bundled by default. LangGraph is an optional extra — only the CLI graph-execution path needs it: pip install 'ai-team-os[langgraph]'.

Verify Installation

# Check OS health (API must be running — port may vary, check api_port.txt)
curl http://localhost:8000/api/health
# Expected: {"status": "ok"}

# Create your first team via CC
# Type in Claude Code:
# "Create a web development team with a frontend dev, backend dev, and QA engineer"

First Words to Your Session

Hooks and MCP tools activate automatically, but a fresh model treats them as background noise until you make them the working protocol. Open your first session after install with one sentence:

"This project runs on AI Team OS - learn its tools and rules (start with /os-help), and use its task wall, memos and memory for everything you do."

One sentence is enough - the hooks keep the session honest from there.

Tool Loading Configuration (optional)

By default the MCP server registers all 113 tools. Two startup environment variables let you trim the surface for leaner sessions or non-CC clients with tool-count limits (e.g. Cursor only forwards the first 40 tools). Both are read once at server startup - no runtime state, no restart-on-change.

AITEAM_TOOLSETS - pick which capability-domain groups register:

  • unset or all - full 113 (backward compatible)
  • default - core groups only (task,team,memory,infra,reports = 29 tools, hard-capped at <=50)
  • a comma list of group names, mixable with default for incremental loading, e.g. AITEAM_TOOLSETS=default,ecosystem
  • unknown names are warned on stderr and ignored (a config typo never blocks server start)

AITEAM_READONLY=1 - orthogonal overlay that strips every write tool (create/update/delete/apply/send/... plus os_restart_api) after registration, keeping only read tools. Handy for audit/observer sessions.

The 16 groups (default groups marked *):

Group Tools Group Tools Group Tools
task * 8 project 6 links 3
team * 5 agent 7 channels 3
memory * 6 meeting 10 task_analysis 2
infra * 7 briefing 4 watchdog 1
reports * 3 analytics 2 workflows 3
ecosystem 42 
# Example: lean core + ecosystem, read-only
AITEAM_TOOLSETS=default,ecosystem AITEAM_READONLY=1 <launch CC / MCP server>

Uninstall

# Plugin install:
claude plugin uninstall ai-team-os
# Then manually remove residual data:
# Windows: rmdir /s %USERPROFILE%\.claude\plugins\data\ai-team-os-ai-team-os
# Unix:    rm -rf ~/.claude/plugins/data/ai-team-os-*
# Restart Claude Code to stop active hooks.

# Source install — full cleanup:
python scripts/uninstall.py
# Preview first:
python scripts/uninstall.py --dry-run

Start the Dashboard (optional)

cd dashboard
npm install
npm run dev
# Visit http://localhost:5173

Dashboard Screenshots

Command Center

Command Center

Team Working — Live Activity Tracking

Team Working

Task Board

Task Board

Workflows — CC ultracode Run Observability

Persistent governance layer for CC ultracode Workflow runs — every run is auto-tracked as a team, surfacing stage progress plus per-agent token and tool-call telemetry. Workflows

Workflow Detail — Phase Swim Lane & Per-Agent Telemetry

Drill into a single run: a phase swim lane aligns every stage against one timeline, and a per-agent telemetry table breaks down tokens, tool calls, duration and state per stage — with a failed contract check surfaced in red. Workflow Detail

Project Detail — Decision Timeline

Decision Timeline

Project Detail — Leader Context & Worktrees

Per-project roll-up of every attached Leader session with live context watermarks, alongside the Git worktrees in play — flagging any that carry uncommitted changes. Project Detail

Agent Board — Live Agent Lanes

Real-time roster of every agent across teams — busy / waiting / offline tallies up top, each card showing its current task, context watermark and last-active time, blending workflow agents with named specialists and their historical trails. Agent Board

Meeting Room

Meeting Room

Ecosystem Research Platform

The ecosystem archive's initial listing — the full set of tracked open-source repositories with stars, primary language and topic tags, ready to open into per-repo research and integration. Ecosystem

Activity Analytics

Analytics

Event Log

Events

Auto-Wake System — Autonomous Task Advancement

Auto-Wake Demo

Auto-Wake System

The Leader supports scheduled auto-wake to autonomously advance tasks without supervision:

  • Automatically checks context usage and pending tasks every 10 minutes
  • When tasks are available, autonomously creates teams and assigns work
  • When user decisions are needed, records them asynchronously via the Briefing system
  • When context exceeds 80%, auto-saves progress and prompts to open a new session

Ecosystem Integration Recipes

AI Team OS is designed as a meta-plugin — it orchestrates other MCP servers rather than reimplementing their capabilities. Pre-built recipes let you integrate popular tools in minutes:

Recipe Integrates With What You Get
GitHub @modelcontextprotocol/github Auto PR creation, issue tracking, code review coordination
Slack @anthropics/slack-mcp Team notifications, decision escalation, status broad