Tencent's TeamAI: Git-Native Sync for Team AI Agents Across 20 Tools

Tencent recently open-sourced TeamAI, a command-line tool that addresses a growing pain point: teams using different AI coding agents (Cursor, Codex, Claude Code, etc.) end up with fragmented configurations, causing inconsistency and onboarding friction. TeamAI solves this by treating AI assets as code — storing them in a shared git repository (GitHub, GitLab, GitCode, CNB, TGit, or private Git) and syncing them automatically via familiar git workflows.

What It Is

TeamAI is a CLI that lets a team place 12 categories of AI assets — skills, rules, env, agents, hooks, MCP, models, learnings, codebase graph, teamwiki, usage, sessions — into a single git repo. An admin initializes the repo with teamai init https://github.com/your-org/your-repo; members run the same command once (optionally with --scope user for user-level install). Thereafter, every AI session automatically pulls the latest assets — no manual sync needed.

Unlike personal skill libraries (e.g., taste-skill, diagram-design) that make your agent stronger, TeamAI makes the team's agents consistent. The project's README includes an eight-scenario before/after comparison; the most striking scenarios are:

New-hire onboarding : previously "where to start? too many docs"; now team skills, project context, tools, and MCP configured in one go.

Unified output & standards : product specs, code style, architecture guidelines, copy standards — all AI output aligns to the same baseline.

Reusing real experience : solved problems, discussions, code references become part of the agent's knowledge.

Switch any agent, keep team skills : changing AI tools doesn't require reconfiguration; all agents see the same team skill set.

By moving asset collaboration onto git, teams gain branch isolation, PR review, main-branch protection, and change history — twenty years of software-engineering governance applied to AI assets. A broken skill can be rolled back; authorship is traceable — far better than passing latest-v3-final.zip in chat groups.

Three-Layer Capability Model (Official Product Overview)

Team Execution (stable): lets agents work the team way. Covers seven asset types — skills, rules, env, agents, hooks, MCP, models — covering the full agent runtime configuration surface.

Team Context (beta): lets agents understand the team. Includes learnings (captured pitfalls), codebase graph (repository knowledge graph), teamwiki (team knowledge base). This layer bridges assets to organizational memory.

Team Improvement (beta): lets each execution feed back into the team. usage and sessions track which skills are high-frequency and which rules are ignored, enabling data-driven iteration.

Only the first layer is production-ready; the latter two carry explicit beta labels — a point the author emphasizes before committing organizational memory to them.

Compatibility Matrix: 20 Agents, Not "12+"

The author counted the official README's compatibility table row by row: 20 agents are listed, not the vague "12+" often used in marketing. Supported agents include Claude Code, Codex, Cursor, CodeBuddy, WorkBuddy, GitHub Copilot CLI, OpenCode, Pi, Oh My Pi, OpenClaw, Hermes, DeepSeek Harness, Qoder, Kiro, Trae, ZCode, JoyCode, Muse Code, and others — covering most mainstream AI coding agents domestically and internationally. Nine agents (CodeBuddy, WorkBuddy, etc.) support all 12 asset types. The matrix also honestly notes gaps: OpenClaw and Hermes lack agents and MCP sync; Cursor and Copilot CLI lack model sync.

Installation & Configuration

Three roles, three commands:

# Admin: initialize shared repo after creation
npm install -g teamai-cli
teamai init https://github.com/your-org/your-repo

# Member: workspace-level (default) or user-level
teamai init https://github.com/your-org/your-repo
teamai init https://github.com/your-org/your-repo --scope user

Teams that don't want to start from scratch can fork a pre-configured template from the teamai-hub GitHub organization, which includes production-grade skills, rules, and review agents. For Chinese users with GitHub access issues, the Chinese README points to skillhub.cn as an alternative install source. Once installed, all interactions happen via the /teamai chat command: bootstrap, join team, share skill, open dashboard — all in one line.

Drawbacks & Caveats

License confusion : GitHub metadata shows NOASSERTION, causing some toolchains to flag an unknown license. The author verified the LICENSE file: it is standard MIT — commercial use is fine; the issue is GitHub's license detection, not the license itself.

Beta layers : Team Context and Team Improvement (learnings, codebase graph, teamwiki, usage, sessions) are not yet stable. The author advises small-scale pilots before betting organizational memory on them.

Modest traction : 5.5k stars places it mid-tier in this category; ecosystem depth and community cases are still thin. 51 open issues show normal responsiveness, but long-term maintenance remains unproven compared to top-tier projects.

Git is not a free lunch : putting assets in git introduces governance overhead — who reviews PRs, how to prioritize rules, who cleans up stale skills. The tool doesn't solve these; the team must establish its own conventions.

Author's Perspective

TeamAI essentially brings DevOps to AI assets: configuration management, version control, and distribution — proven engineering practices — applied to agent assets. The direction is sound: as team agent count grows from one to ten and assets expand from a few rules to a full system, manual sync and file sharing inevitably break down. Git is the most mature refactoring vehicle today.

Strengths: no vendor lock-in (works with any agent, any git host, including domestic platforms GitCode, CNB, TGit), low barrier for Chinese teams, multilingual README. However, the only solid layer today is Team Execution; alternatives like per-agent config sharing or dotfiles-style scripts partially address the same problem. TeamAI's edge is breadth and convenience, not uniqueness. The real differentiator would be the beta layers — if learnings and teamwiki stabilize, TeamAI could become an OS-level entry point for team AI. Until then, treat it as a capable asset-sync tool, not a strategic bet.

Another observation: the team-collaboration layer is relatively open compared to the red ocean of personal agent enhancers. A company of Tencent's scale investing in infrastructure reduces abandonment risk — a key selection criterion for a tool that must run long-term in team workflows.

Adoption depends on AI usage depth: teams of 3-5 using agents casually don't need it; teams of 10+ with agents deeply embedded in development flow should try — initialization is one command, exit cost is low.

GitHub repository card: https://opengraph.githubassets.com/1/Tencent/teamai-cli