AI Agent Team: Role-Based Slash Commands for Claude Code, Plus a SQLite Thread Memory
AI Agent Team-拥有24/7专业AI开发团队：产品经理、前端开发、后端开发、测试工程师、DevOps工程师、技术负责人。一键安装，支持中英文命令，大幅提升开发效率！
At a glance
- What is it?
- ai-agent-team is an MIT-licensed npm package that installs a set of role agents (PM, frontend, backend, QA, DevOps, tech lead) as Claude Code skills and adds a local MCP server for persistent task threads. The interesting part is the thread layer, not the personas.
- Who is it for?
- Adopt it if you already run Claude Code daily, want named slash commands for role work, and accept that your conversation history will live in a local SQLite file behind an MCP server you have to register by absolute path. Do not adopt it if you expect a multi-agent orchestration runtime: the README describes personas and a memory layer, not agents that call each other.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 78 days ago.
- What is it written in?
- Mainly JavaScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The problem is context loss, not a shortage of prompts
Most prompt packs for coding assistants solve the wrong problem. They give you a better paragraph to paste. The failure that actually costs time is that a session ends, the terminal closes, and the reasoning behind a design decision is gone. ai-agent-team attacks that second problem. The README frames the package as a 24/7 team of roles (product manager, frontend, backend, QA, DevOps, tech lead) reachable through slash commands such as /pm, /fe, /be, /qa, /ops and /tl, and then spends most of its documentation on Thread Manager, described as an AI memory system with semantic search, task thread management and automatic Git version control. The audience is narrow and identifiable: people already running Claude Code who want their project conversations to survive a restart. If you do not use Claude Code, the core feature has nothing to attach to.
Role skills are separate repositories, not one bundle
The README lists the role agents as independently published repos, each installed on its own with a skills command, for example npx skills install peterfei/ai-agent-tech-leader or npx skills install peterfei/ai-agent-qa-engineer. That is a deliberate packaging choice with a practical consequence: you can pull in the QA skill without the DevOps skill, and each role repo can version on its own schedule. It also means the top-level package is closer to an installer and a memory server than a monolith. The README claims compatibility with Claude Code, Cursor, Codex CLI, Gemini CLI and Windsurf, among 50+ runtimes, though the documented quick-start path assumes Claude Code specifically. Treat the wider runtime list as a claim to verify per editor rather than a guarantee, because the installation and verification steps shown are Claude Code commands.
Thread Manager: SQLite plus a bundled embedding model behind MCP
The architecture diagram in the README is the most informative artifact in the repository. A user query reaches the Claude assistant, which calls the Thread Manager MCP server as a tool. Inside that server, a Thread Manager component parses the request and routes it two ways: new messages go to a Messages DAO, and searches go to a vector search engine. The DAO passes text through an embedding service that loads a locally bundled model identified as Xenova/all-MiniLM-L6-v2, and both the messages and the search index land in a SQLite database. Search results come back as a list of related messages, which the Thread Manager formats into context and hands to the assistant. Two design decisions stand out. The embedding model is local, so semantic search does not require an API key or a network round trip. And storage is a single SQLite file, which makes the memory portable and inspectable but also means the quality of your recall depends on a small sentence-transformer rather than a hosted embedding endpoint. The README does not document the database path or schema, so you will have to find the file yourself if you want to audit it.
Installation has a step that silently gates the main feature
The documented path is three commands. First, npm install -g ai-agent-team. Second, ai-agent-team init, which the README marks as mandatory, either globally (affecting ~/.claude/) or per project via cd your-project && ai-agent-team init (affecting ./.claude/). Third, and this is the one people skip, register the MCP server: claude mcp add thread-manager node "/your/path/.claude/skills/thread-manager/dist/index.js". The README states plainly that without step three, /threads and /pm-start will not work. The path is absolute and machine-specific, so it does not survive being copied between machines or into a container image without editing. Verification is manual: run /threads to list threads, /pm-start "测试" to create a test thread, /thread info to inspect the current one. When global and project-local configuration coexist, the README says Claude Code prefers ./.claude/ and falls back to ~/.claude/.
The performance table is marketing, and the Git integration is the real risk
The README includes a comparison table claiming 200%+ efficiency, 10x+ multitasking and infinite context retention versus native Claude. Those numbers have no stated methodology and should be read as promotion, not measurement. The more concrete claim is automatic Git integration: the thread layer creates task branches, tracks file changes and produces code statistics. That is also where the design deserves scrutiny. Automatic branch creation inside an assistant session can collide with whatever branching convention a team already enforces, and the README does not describe how conflicts are resolved, whether branches are named predictably, or how to disable the behaviour for a repository where the assistant should never touch Git state. If you work in a monorepo with protected branches or a strict commit policy, test this on a scratch repository before pointing it at anything that matters. The README also does not state where thread data goes on uninstall, so plan for a leftover SQLite file and a registered MCP entry you will need to remove by hand.
Bundled extras: changelog, SoftCopyright, desktop tidy, note images
Beyond the roles and threads, the package ships several unrelated skills. Changelog Generator analyzes Git history to produce CHANGELOG.md, with a documented command set: changelog-generate generate --all --format html, changelog-generate release --github-release, and changelog-generate update for incremental runs. SoftCopyright scans a project and emits software copyright application material, described as a manual and source listing at 50 lines per page up to 60 pages, with comment stripping and PDF export via the browser print dialog, across 20+ languages. TidyMyDesktop classifies desktop files, detects version numbers and keeps the newest, with a dry-run preview. DrawNote turns notes into styled images. These are conveniences bundled under one installer, and their presence explains why the repository topics include notebook and drawnotes. They are not the reason to adopt the package, and each one is a small surface with its own behaviour to learn.
Alternatives and the honest boundary
The closest comparison is the assistant's own built-in memory or a plain context file checked into the repository. A CLAUDE.md or equivalent notes file is deterministic, diffable, reviewable in a pull request and free of any server process. Thread Manager trades that transparency for semantic recall over accumulated history and the ability to switch between parallel task threads with /thread switch <id>. If your work is one long-running project where decisions are few and written down, the file wins. If you juggle several feature threads and keep losing the reasoning behind earlier choices, the search layer earns its keep. What this project is not, based on the material, is an orchestration framework: there is no documented mechanism for one role agent to invoke another, hand off work, or run autonomously in the background despite the 24/7 framing. The roles are entry points and the threads are memory. Anyone expecting agents that coordinate with each other will be disappointed by what the README actually describes.
Licence and upkeep
The package is MIT-licensed, which permits commercial use, modification and redistribution provided the copyright notice and permission notice are retained. That is a permissive arrangement and imposes no copyleft obligation on your own code. It says nothing about the separate role skill repositories, each of which carries its own licence that you should check individually before bundling them into a product. On maintenance: the release history shows v2.0.0 in December 2025, v2.0.1 shortly after, and v2.1.0 in June 2026, so the project is active on a roughly six-month cadence at the top level. The README still advertises version 2.0.0 in its badge while the latest release is 2.1.0, a small sign that documentation lags releases. The bundled embedding model and the MCP server binary path are the two things most likely to break on upgrade, since the registration command hardcodes a path into the installed skill directory.
Editorial conclusion
Adopt it if you already run Claude Code daily, want named slash commands for role work, and accept that your conversation history will live in a local SQLite file behind an MCP server you have to register by absolute path. Do not adopt it if you expect a multi-agent orchestration runtime: the README describes personas and a memory layer, not agents that call each other. Before trusting it, run /threads and /thread info after install, then inspect the SQLite database the MCP server created to confirm what is actually persisted.
Community notes