Tencent Cloud TeamAI was formally open-sourced on October 10, 2026. The team AI collaboration tool, long used internally at Tencent, puts skills, work rules, tool configurations and project knowledge into a single Git repository, where they are reviewed, updated and versioned like code, then synced to the agent each member uses. It already works with 16 agents, including WorkBuddy, CodeBuddy, Claude Code, Codex and Cursor, and its official repository sits on GitHub under the Tencent organization as teamai-cli.
Why Git, rather than yet another admin console
The first problem teams hit with agents is usually not model quality but scattered configuration: one person updates a rule, another still runs the old version, and a useful skill lives on a single colleague's machine. TeamAI treats these assets as code. When a member submits a new skill or changes a rule, the system opens a branch and a merge request for team review; once merged, agents that support session-start hooks pull the update automatically when a new conversation begins. Admins can also push model configurations centrally, assigning approved model services to tasks involving internal data and knowledge.
Turning experience into a searchable team asset
Its handling of failure is just as notable. When a member repeatedly corrects an agent, or tool calls keep failing, TeamAI prompts them at the end of the session to share what happened; the agent writes the problem and the fix into a document stored back in the team repository, ready to be retrieved on the next similar task. TeamAI is also integrated into Tencent Cloud's ClawPro agent management platform, where admins manage skills and MCP configurations centrally and distribute them to connected agents.
Read the internal numbers with their scope in mind
Tencent Cloud has shared one internal test: across 13 R&D tasks, a lower-cost model with access to team knowledge scored 6.1% higher overall than a higher-tier model without it, while costing 76% less to run. Thirteen tasks is a small, vendor-reported sample, not a universal law. But the direction is concrete: when team knowledge is structured and reliably distributed, adding knowledge can pay off more than simply buying a more expensive model. For teams already coding with agents together, TeamAI is less about whether agents work, and more about whether hard-won experience stays in the team.
