Tencent's TeamAI CLI added 1,215 GitHub stars in the one-day snapshot supplied to MrKeyoor on September 8, the largest daily gain in this brief. Developers arriving for its DeepSeek Harness support should read the smaller print in Tencent's compatibility matrix: five of 13 capability columns are checked for that client. Skills, documentation, learnings, codebase data and teamwiki are covered. Rules, environment settings, subagents, hooks, MCP configuration, usage data, sessions and the dashboard are absent.
The repository's scope is broader than its DeepSeek adapter. TeamAI's package manifest describes an MIT-licensed TypeScript command-line tool that distributes team skills and maintains a shared knowledge base through Git. The README currently shows all 13 checks for Claude Code, Codex, Cursor, CodeBuddy and Qoder. WorkBuddy, OpenCode, OpenClaw, Hermes and DeepSeek Harness each get a smaller subset.
A Git workflow for agent behavior
TeamAI puts agent-facing material in a shared repository: skills, rules, documents, hooks, MCP definitions and team learnings. A contributor runs teamai push, the CLI opens a branch and merge request, and a reviewer can inspect the change before it reaches colleagues. After merge, teamai pull distributes the approved files into each supported tool's native directories. The README's workflow uses the existing code-review habit as the control point for changing how coding agents work.
The setup has more machinery than copying a folder. According to the usage guide, TeamAI requires Node.js 20 or newer and Git. Project scope is the default. In that mode, the tool keeps its repository clone, state, search index and cached resources under a per-project partition in ~/.teamai/projects/, while agent-specific directories such as .claude or .codex live in the working project. User scope is available for conventions that should follow a developer across projects.
The documented setup starts with two commands, followed by interactive choices for repository access and target agents:
npm install -g teamai-cli
teamai init https://github.com/yourorg/your-team-repo
Teams can narrow what each person receives. The distribution controls map roles to namespaces, filter skills and rules by tags, and subscribe to additional TeamAI-formatted repositories. As a shared store grows beyond a few instruction files, those filters prevent every session from receiving the full collection. A mobile developer does not need an infrastructure agent's deployment rules in every session, and the repository's merge history records who changed each resource.
DeepSeek Harness gets the knowledge layer
The project's three-layer model separates execution, context and team improvement. DeepSeek Harness receives the parts centered on reusable knowledge: skills and docs on the execution side, then learnings, codebase data and teamwiki. The support row leaves its rules, env, agents, hooks and MCP cells empty. Its usage, session and dashboard cells are empty as well. TeamAI can therefore give a DeepSeek Harness session shared instructions and project memory. Its README does not promise the policy injection or reporting available to the five fully checked clients.
DeepSeek Harness has an explicit target ID. The single-repository setup documentation assigns it the dsh identifier, so an administrator can select it explicitly with --agent dsh. In this mode, knowledge lives under .teamai on the business repository's main branch, while reports use a separate orphan branch and machine-local tokens remain ignored by Git. A clone can initialize the team configuration when the developer already has provider authentication.
Automatic delivery deserves a test in each DeepSeek environment. TeamAI's general flow triggers teamai pull from a SessionStart hook, yet the DeepSeek Harness row has no hook check. Tencent's guide documents the dsh selector but does not show a DeepSeek lifecycle adapter that would fire that pull. The documentation establishes file compatibility. It does not establish session-start synchronization. Teams should verify update behavior with a throwaway repository before they depend on it.
The local trust boundary grows
A reviewed Git repository can still distribute executable behavior. TeamAI lets administrators declare shell commands in hooks/hooks.yaml, and teamai pull can install them in supported clients. Its hook controls include an autoApply: false mode that asks for manual confirmation, a restriction that accepts commands only from the team's scripts directory, and a local variable that disables team hooks. Codex adds its own trust step for new or changed hooks, which TeamAI says it does not bypass.
MCP credentials create a separate local risk. The guide's secrets section tells administrators to place ${VAR} placeholders in the shared YAML. TeamAI resolves each placeholder before writing a client's configuration, so the resulting token is stored as literal text on disk. New files receive 0600 permissions, but project configurations such as .mcp.json, .codex/config.toml and opencode.json must also be ignored by Git and kept out of commits.
Single-repository mode has another plain-text boundary. The same guide says .teamai/env/env.yaml is committed to the main branch and travels with every clone. Tencent explicitly limits that file to non-secret shared configuration. An organization adopting TeamAI therefore needs two reviews: one for what agents are instructed to do, and another for which generated files or local settings may hold credentials.
MrKeyoor tested the code outside Tencent's own CI. Our review of TeamAI CLI covers the setup reality: commit cb51ebc built in eight seconds in a clean Debian container, all 2,729 project tests passed, and npm audit reported no known dependency vulnerabilities during that run. The install pulled 360 packages and occupied 895 MB, which is worth budgeting for CI images and managed developer machines. Those results cover that tested commit and dependency snapshot, not every future release.
Memory collection has explicit switches
TeamAI's experience-sharing feature watches for friction such as a user interrupt, a rejected tool call or repeated command failures. When a session crosses its threshold, the README says the agent suggests running a share-learnings skill. It does not automatically publish the lesson. Automatic knowledge recall is also off by default and can be enabled or disabled at team and member level. Those defaults give teams a chance to inspect the behavior before shared session history starts influencing later work.
The reporting path is narrower than a transcript archive. Tencent's privacy notes say intervention tracking stores counts rather than prompt or transcript text. A session pushed to the team repository contains counts and tool names by default. Including the first request takes an explicit --include-prompt flag, after which the CLI applies its secret redaction. Local monthly logs keep that redacted first line and are pruned after 90 days. Operators still need to decide whether tool names, counts and optional prompts fit their own data policy.
For code recall, teamai import builds a graph of components, interfaces and import edges. The implementation description uses a WebAssembly tree-sitter parser for TypeScript, JavaScript, Python and Go. Other languages, including Java and Rust, use regex-based extraction. If the parser cannot load, TeamAI records an AST_UNAVAILABLE gap and falls back to the heuristic path. That distinction should stay visible when teams judge the quality of a recalled dependency link.
Forty-eight releases before 1.0
The npm registry listed version 0.23.0 on September 8, the 48th published version since the package appeared on May 1. The v0 label says little about code quality. It does warn consumers that configuration paths and defaults can still move quickly. Pinning the CLI in CI and reading its migration notes is more defensible than letting every workstation update on a different schedule.
One current gap sits outside the agent matrix. In open issue 416, a user reports that skills copied from ordinary public repositories retain no upstream Git relationship, so TeamAI cannot tell when those copies become stale. The built-in subscription path follows updates, according to the report, but expects the upstream repository to use TeamAI's own layout and metadata. The proposed provenance and refresh commands have no linked implementation or assigned owner on the issue at the time of writing.
Watch whether Tencent fills more of the DeepSeek Harness support row and documents a session-start sync path for dsh. Teams piloting 0.23.0 should test hook approval and inspect every generated MCP file before widening access. A 1,215-star day measures attention. Repeatable adoption depends on stable update paths and predictable files on disk.