Tencent's teamai-cli Treats Your Team's Taste as a Git Repo
teamai-cli picked up 563 stars in a day and its pitch fits the problem most teams hit around the third month of agent adoption. Everyone on the team uses a different agent. Claude Code, Codex, Cursor, CodeBuddy, OpenCode. Each one carries its own rules, its own skills, its own idea of how your codebase works. Nobody's context reaches anybody else. Tencent's answer is to make the team's shared knowledge a git-managed artifact that every agent reads.
The mechanism is deliberately unexciting: skills, rules, docs, environment variables, and hooks are distributed through push, review, merge, pull. Your team's coding standards get code review. That is the whole design decision, and it is the right one, because the alternative everyone is living with is a CLAUDE.md that one person edits and nobody else knows changed.
Two beta pieces are more ambitious. Team Context captures what it calls friction-based learnings from coding sessions, meaning the moments where an agent got stuck, and makes them recallable in later sessions by other people. Team Improvement adds usage analytics and session tracking to keep the knowledge base from rotting, which is the failure mode nobody talks about: a skills directory nobody prunes becomes a liability inside a quarter. MIT licensed, 695 commits, active as of yesterday.
The skills shelf has been consolidating for months, from personal repos (https://clauday.com/article/11c3de84-b771-435a-b319-0713947bcd1d) to entire GitHub trending pages (https://clauday.com/article/e3d92bdf-4f28-4492-9d9d-dce2e7ffcb3a). This is the corporate turn of the same wheel: once skills matter, somebody has to own their release process. https://github.com/Tencent/teamai-cli
← Back to all articles
The mechanism is deliberately unexciting: skills, rules, docs, environment variables, and hooks are distributed through push, review, merge, pull. Your team's coding standards get code review. That is the whole design decision, and it is the right one, because the alternative everyone is living with is a CLAUDE.md that one person edits and nobody else knows changed.
Two beta pieces are more ambitious. Team Context captures what it calls friction-based learnings from coding sessions, meaning the moments where an agent got stuck, and makes them recallable in later sessions by other people. Team Improvement adds usage analytics and session tracking to keep the knowledge base from rotting, which is the failure mode nobody talks about: a skills directory nobody prunes becomes a liability inside a quarter. MIT licensed, 695 commits, active as of yesterday.
The skills shelf has been consolidating for months, from personal repos (https://clauday.com/article/11c3de84-b771-435a-b319-0713947bcd1d) to entire GitHub trending pages (https://clauday.com/article/e3d92bdf-4f28-4492-9d9d-dce2e7ffcb3a). This is the corporate turn of the same wheel: once skills matter, somebody has to own their release process. https://github.com/Tencent/teamai-cli
Comments