AI Team OS
Give Claude Code and Codex a shared record of tasks, memory, and messages so work can continue across sessions.
- Source repo
- CronusL-1141/AI-company
- Stars
- ★ 373
- Last updated
- 12d ago
- License
- MIT
- Primary language
- Python
- FA score
- 81/100 · Good
At a glance
- How it runs
- Works with
- Portable with changesClaude CodeCodex (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A Claude Code development team needs task ownership, progress memos, and decisions to persist across sessions.
- Not a fit if
- Teams that need Codex to wake and continue after a turn ends
- Teams requiring complete Codex token and cost attribution
- Teams that do not use Claude Code or Codex
- Source review
- 81/100 · Good
What does this agent do, and when should you use it?
AI Team OS is a local collaboration layer for Claude Code and Codex, built from Python 3.11+, FastAPI, SQLite, an MCP server, and a React 19 Dashboard. Each host connects through its own adapter, configuration, and native agent tools; Claude Code can also install lifecycle hooks and templates. The OS stores tasks, progress memos, project memory, reports, channel messages, and activity observations for later authorized sessions to retrieve. Agent reasoning and execution remain with Claude Code or Codex sessions, and persistent records do not imply a resident model or automatic wakeup after a Codex turn ends. It fits development teams that need traceable handoffs across sessions, with setup work to configure host adapters and run the local API.
Once installed and running, Claude Code or Codex clients use MCP tools to read and write project tasks, task_memos, direction memory, reports, and channel messages; the REST API provides the shared backend and the React Dashboard displays tasks, activity, projects, and reports. A session can call context_resolve for project context, use task_create to create tasks with parent or dependency links, record progress with task_memo_add, then hand off messages with channel_send, channel_read, and channel_wait. The Claude Code adapter can also record session and tool activity through hooks and provides Claude-specific context injection and Workflow observation. The system offers local BM25 retrieval through memory_search and combined memo, report, and task search through unified_search; these retrieval operations do not call a model. Actual agent execution depends on each host's native tools and permissions; the OS produces durable records, activity observations, and Dashboard views.
- A Claude Code development team needs task ownership, progress memos, and decisions to persist across sessions.
- A project using both Claude Code and Codex needs shared channels for handoffs while keeping each host's identity and configuration distinct.
- A project lead wants to inspect current work, blockers, and activity history through the task wall and Dashboard.
- A researcher wants to turn repository analysis into reports and searchable memory for later project sessions.
- An engineering team needs a structured discussion record and can use meeting_create to select a template and save its conclusion.
How do you install or deploy this agent?
Requires Python 3.11+, uv, and Claude Code or Codex with MCP support. The optional Dashboard frontend requires Node.js 20+. Install into system Python rather than a project virtual environment. For Claude Code, choose either the plugin or source installation path, not both. Codex has a separate adapter path; the Claude Code installer does not configure it.
How do you use this agent?
Claude Code plugin installation:
pip install uv
claude plugin marketplace add CronusL-1141/AI-company
claude plugin install ai-team-osClaude Code source installation (choose one path):
git clone https://github.com/CronusL-1141/AI-company.git
cd AI-company
python3 install.pyCodex adapter installation:
git clone https://github.com/CronusL-1141/AI-company.git
cd AI-company
python3 -m pip install -e .
python3 scripts/codex_adapter.py install --api-url http://127.0.0.1:8000
python3 scripts/codex_adapter.py statusReload the relevant host and review Codex Hook trust prompts. After confirming the API is available, call context_resolve in a session, read the project tasks and memos, and use the MCP tools to record work. Health check:
curl http://localhost:8000/api/healthOptional Dashboard development launch:
cd dashboard
npm install
npm run devWhat are this agent's strengths and limitations?
- One project record for tasks, memory, reports, and messages is available to Claude Code and Codex sessions while preserving their distinct native identities and settings.
- SQLite persistence and local BM25 search provide durable records and retrieval without a separate model call for search.
- Task parent-child links, dependency unlocking, explicit waits, and progress memos support recorded handoffs.
- The React Dashboard presents tasks, activity, reports, and Claude Code Workflow observations.
- Claude Code and Codex require separate adapter setup; the Codex path includes source installation, dependencies, and Hook trust review.
- The OS does not run a resident model, and Codex does not wake automatically when a message arrives after a turn ends.
- Complete Codex token and cost attribution is not available; usage without evidence remains unknown.
- The README says Codex's on-demand runtime was validated on macOS but has not been validated on Windows.
How does this agent compare with similar options?
The README compares AI Team OS with CrewAI, AutoGen, LangGraph, and Devin, positioning it as a shared task, memory, and observation layer for native Claude Code and Codex agents rather than a standalone agent framework or hosted AI engineer.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| AI Team OS This agent | 81 · Good | Agent plugin / skillFree + model costs | ★ 373 | 12d ago | Python | Claude Code |
| AgentRQ | 64 · Some gaps | Self-hosted serviceFree + model costs | ★ 1.2k | today | Go | Codex · Claude Code |
| COMPASS Skills | 85 · Good | CLIFree | ★ 761 | 4d ago | Python | Codex · Claude Code |
| Cindy | 52 · Major gaps | Desktop appFreemium | ★ 3.4k | today | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The materials describe read-only tool profiles, restrictions on selected destructive tools, a localhost API, an explicit external-execution entry point, allowlisted receipts, and controls such as atomic writes, backups, CAS, rebuildable summaries, and no-write failure paths. These support strong data-boundary and recovery practices. Deductions: restrictions are shown for only some templates, so least privilege across all tools is not established; external calls can still execute and incur costs, while confirmation mainly relies on host controls; repository identity and a vulnerability-reporting path are present, but the publisher is unverified and response is best-effort.
The documents detail arithmetic consistency across paths, rebuild protocols, concurrency guards, and failure semantics; CI configuration lists tests and invariant checks. Deductions: static materials cannot establish that tests passed; README product figures differ from the supplied description (108/40+ versus 118/25), weakening consistency; service availability also depends on third-party packages without an availability guarantee.
The intended users, Claude Code/Codex scenarios, read-only audit, research, and task coordination uses are specific. Host differences, unsupported automatic wake behavior, unknown identity, frozen features, and Codex usage-attribution gaps are clearly bounded. Deductions: the documentation is oriented mainly toward developers and named hosts, and some features depend on host capabilities and local configuration.
The README is organized by capability and links to usage documents and a changelog; the MIT license is explicit, and version history and security policy describe supported versions, reporting, and responsibility. Deductions: the supplied materials omit the installation guide body, so complete setup steps cannot be verified; README counts conflict with the candidate description, reducing naming/count stability; the security policy says only the latest release line is supported but lists 1.11.x, which does not align cleanly with the stated 1.15.2 release.
The task wall, cross-session records, search, reports, and dashboard provide reusable coordination outputs, and the materials distinguish persistent records from a continuously running model. Deductions: value depends on users configuring local services, host tools, and integrations; the dependency and maintenance burden is nontrivial, and the materials do not quantify overall user cost versus benefit.
Multiple design documents specify mechanisms, boundaries, failure cases, and named tests; CI configuration lists automated checks, and the documentation distinguishes recorded facts, unknowns, and estimates. Deductions: this review has only the supplied static materials, with no result logs or independent corroboration, and headline counts conflict with the detailed README.
- The README's 118 MCP tools and 25 templates differ from the candidate description's 108 tools and 40+ templates; verify the feature inventory for this revision before adoption.
- The security policy lists 1.11.x as supported while the README identifies 1.15.2; confirm the current security-support scope and installation steps.