TencentTeamAI keeps a team's skills, rules, docs, hooks, MCP servers, and environment variables in a shared git repository and distributes them to each member's local AI coding tools. Coverage spans Claude Code, Codex, Cursor, CodeBuddy, WorkBuddy, OpenCode, OpenClaw, Hermes, and DeepSeek Harness, with a documented per-agent matrix of which capabilities each one receives, and the shared repository can live on GitHub, GitLab, GitCode, CNB, TGit, or a private git service.
Distribution follows a push, review, pull flow. teamai push writes a change to a branch and opens a merge request — re-pushing an unmerged resource updates the existing request instead of opening a second one — and once it is merged, teamai pull syncs the result into each tool's directories. A SessionStart hook runs the pull automatically, so a session begins with the current resources. Installation is scoped either to a project directory or to the user's home directory.
Administrators shape what each member receives. Roles map members to namespaces so they sync only the skills for their role, tags let members subscribe to a subset of skills and rules, and sources subscribe the team to additional skill repositories from other teams or elsewhere in the organisation. Hooks are declared once in hooks/hooks.yaml and MCP servers once in mcp/mcp.yaml, with ${VAR} placeholders for secrets, and each pull writes them into every installed tool's native configuration.
TeamAI also builds a searchable knowledge base. teamai recall searches accumulated team knowledge with BM25 scoring boosted by a codebase graph, exposed to agents through a teamai-recall subagent that runs a relevance precheck first; it is off by default. teamai import parses repositories into a graph under teamwiki/, using a WebAssembly tree-sitter parser for TypeScript, JavaScript, Python, and Go to resolve imports, call sites, and interface implementations, and a regex track for other languages.
A Stop hook scores each finished session by friction — interruptions, corrections, denied tool calls, and retried failing tools — and suggests documenting it when the score is high enough, at most once per session. teamai digest produces a weekly digest of token usage, conversation volume, and intervention rate, teamai session save records privacy-scrubbed session summaries that feed it, and teamai dashboard shows live session status plus a knowledge-base health page covering coverage, top and silent entries, recall trends, and maintenance.
Features
- Shared harness repository: skills, rules, docs, agents, hooks, MCP servers, and environment variables live in one git repo
- Review-gated distribution: teamai push opens a merge request; a merged change reaches members on the next pull
- Automatic sync: a SessionStart hook runs teamai pull so every session starts from the current resources
- Broad agent coverage: Claude Code, Codex, Cursor, CodeBuddy, WorkBuddy, OpenCode, OpenClaw, Hermes, and DeepSeek Harness, with a documented capability matrix
- Roles and tags: namespace mappings and tag subscriptions narrow what each member syncs
- Skill sources: subscribe to other teams' or the organisation's shared skill repositories and browse them from the CLI
- Declarative hooks and MCP: one YAML file each, reconciled into every installed tool's native config
- Team packages: npm packages and Claude Code plugins are declared, shared, and restored with teamai install
- Knowledge recall: BM25 search with graph boosting, delivered through an opt-in subagent that prechecks relevance
- Codebase graph: teamai import builds a component, interface, config, and dependency graph, with an AST track and a regex fallback
- Friction-scored sessions: a Stop hook flags sessions worth documenting based on interruptions, corrections, and tool retries
- Usage insight: weekly digests, privacy-scrubbed session summaries, and a dashboard with a knowledge-base health console
