AI Development Team

A VS Code extension that runs an autonomous development team with the Claude Agent SDK, on your own Claude Code and/or OpenAI Codex installation (see Requirements).

AI Development Team is an independent project. It is not affiliated with or endorsed by Anthropic or OpenAI. "Claude" and "Claude Code" are trademarks of Anthropic. "OpenAI" and "Codex" are trademarks of OpenAI.

The extension ships no agents. You define the team on the chat's Agents page (or with the Agent Team: Manage Agents command): an orchestrator that runs the conversation, and any specialists it can delegate to. Each agent is stored where you choose:

• Project folder — .claude/agents/<name>.md in the workspace (created when you save the first agent), shareable through version control.
• User folder — ~/.claude/agents/<name>.md, available in every workspace.
• Workspace or user settings — the agentTeam.agents setting.

Pick which agent runs the conversation from the agent selector under the chat input. Without one, the chat runs plain Claude Code.

Project instructions and skill references can live in the agentTeam.instructions setting. The orchestrator and every subagent receive them on both Codex and Claude Code, including mixed teams in either direction. Open Agent Team: Edit Project Instructions and save once in Workspace settings for this project, or User settings for every workspace. Both scopes apply.

For example, keep a shared skill in team-skills/workflow/SKILL.md and enter:

Before starting any task, read and follow team-skills/workflow/SKILL.md.
This applies to the orchestrator and every developer and tester.

You do not need to repeat that reference in AGENTS.md and CLAUDE.md. Every agent receives the original workspace path as a fallback for skill files missing from its worktree, including untracked or ignored skills. Agents are instructed to read the skill and its supporting files there without modifying the source checkout. Paths inside a skill, such as references/schema.json, resolve relative to that skill's directory. This also applies when the skill reference comes from a native instruction file and the shared setting is empty; file permissions still apply.

Commit skills and their supporting files when you want reproducible, self-contained worktrees; otherwise the fallback uses the files currently in the original workspace. Use a path available in every target workspace for a user-level instruction. Changes to the setting apply on the next session start or resume and on each subagent's next run, including Codex follow-ups.

These are shared prompt instructions to read a skill file, not native skill installation or slash-command registration. Each runtime still loads its own native instruction files and skills; the extension does not copy AGENTS.md into Claude or CLAUDE.md into Codex. Keep runtime-specific rules there and common rules in Project Instructions.

MCP servers are managed on the MCP servers page (gear menu, or Agent Team: Manage MCP Servers). They are stored in Claude Code's own scopes, so they also work with the claude command:

• Project — .mcp.json in the workspace, shared through version control. Reference secrets as ${ENV_VAR} instead of writing them in.
• Local — this workspace only, private to you.
• User — every workspace, private to you.

Scout

Scout maps the components of the open workspace: classes, interfaces, UI components, and containers (components that render other components). Open it from the radar button in the chat toolbar or with Scout: Open Component Map.

• Languages: TypeScript and JavaScript (React, Angular, Lit), Vue, Svelte, C# and Razor, Java, Kotlin (Jetpack Compose), Scala, Go, Rust, Python, PHP, ColdFusion, Swift (SwiftUI), Dart (Flutter), C++ and Ruby.
• Relations: renders (a container renders a component), extends, implements (including Go's implicit interfaces), and uses (a type references another). Choose which relations shape the map.
• Only the workspace is scanned; dependencies, build output and tests are skipped (agentTeam.scout.exclude). Click a card to open its source.

Security

• The extension only runs in trusted workspaces (VS Code Workspace Trust).
• Before a session loads a project's own Claude Code configuration (hooks, environment variables, permission rules, credential helpers in .claude/settings*.json, and .mcp.json servers), it lists what that configuration would run and asks you. The answer is remembered per workspace until those files change; Agent Team: Review Project's Claude Code Configuration asks again.
• The executable (agentTeam.claudePath, WSL settings) and pre-approved tools (agentTeam.allowedTools) can only be set in your user settings, never by a project.

Requirements

Agent Team runs on command-line tools you install and sign in to yourself. It installs nothing, links to no downloads, and does not sign you in or manage your accounts. You need at least one of the two CLIs below. Install them in the environment where the extension runs: Windows, macOS, Linux, or the WSL distro of a WSL window.

• VS Code 1.138 or later.
• Claude Code, for agents on Claude models, Scout, and the project-configuration review.
  • Install it as its documentation describes, for example npm install -g @anthropic-ai/claude-code.
  • Set it up: run claude and sign in with /login, or configure an API key, Amazon Bedrock or Google Vertex.
  • If claude isn't on your PATH, set agentTeam.claudePath (or agentTeam.wslClaudePath in WSL mode).
• OpenAI Codex CLI, for agents on OpenAI models (gpt-…).
  • Install it from openai/codex, for example npm install -g @openai/codex.
  • Sign in: run codex login with ChatGPT or an OpenAI API key.
  • If codex isn't on your PATH, set agentTeam.codexPath (or agentTeam.wslCodexPath in WSL mode).
• git, for agents set to run in their own git worktree.
• Node.js and npm, only if you install the CLIs with npm.

While neither CLI is ready, the chat view shows a sign-in page instead of the chat. It gives each CLI's status and how to sign in, and its View requirements button opens this README. Agent Team: Check Claude Code Sign-in and Agent Team: Check OpenAI Codex Sign-in check again. Agents whose CLI is not signed in are greyed out.