AI Workspace Template: Multi-agent Super Power for Your Antigravity

amy 10/04/2026

What is this?

Antigravity AI Workspace Template is a portable, open-source AI workspace engine that turns your codebase into a self-aware knowledge system.

It’s not an IDE plugin or a SaaS lock-in, it’s a lightweight architecture (CLI + Python engine) that works with any LLM, any editor, and keeps your data under your control.

What does it do?

  • ag-refresh: Spins up a multi-agent cluster where each module in your project gets its own “ModuleAgent” that reads the actual source code and writes a concise knowledge doc. No more dumping entire repos into context.
  • ag-ask: When you ask “How does auth work?”, a Router Agent finds the right ModuleAgent, which answers using real file paths, line numbers, and git history, grounded, not guessed.
  • .antigravity/ folder: A shared, IDE-agnostic knowledge hub. VS Code, Cursor, Claude Code, or any editor can read rules.md, conventions.md, or CLAUDE.md from this single source of truth.
  • MCP support: Register as an MCP server so tools like Claude Code can call ask_project() directly—no copy-pasting, no context switching.

How does it work?

During ag-refresh, the engine smartly groups related files by imports, directory structure, and naming patterns, then assigns each cluster to a dedicated ModuleAgent that analyzes ~30K tokens of focused source code in a single pass, no tool calls, no build noise.

A RegistryAgent then compiles all module summaries into a semantic map, so when you ag-ask a question, the Router instantly finds and queries the right expert Agent, grounded in real code.

Optionally, GitAgent adds historical context (who changed what and why), and GitNexus unlocks deeper insights like semantic search and impact analysis across the entire codebase.

Why developers actually care:

  • No more style drift: Agents read .antigravity/conventions.md and match your patterns on the first try.
  • Faster onboarding: New repo? Run ag-refresh—ModuleAgents self-learn the architecture so you (and your AI) don’t have to reverse-engineer it.
  • IDE-agnostic consistency: Switch editors? Your rules, docs, and agent knowledge travel with the code in one folder.
  • Fewer hallucinations: Answers are grounded in real source files, not vague summaries. Tested at 9/10 hallucination resistance on a 374-file project.
  • Zero vendor lock-in: Plug in any OpenAI-compatible API (local LLMs included), use any IDE, no proprietary cloud required.
  • Git-aware: A dedicated GitAgent understands who changed what and why—so answers include context, not just code.
  • Privacy-first by design: Runs locally, respects your .env, and keeps sensitive logic out of third-party dashboards.

Bottom line: It makes your AI teammate actually know your codebase, without the context bloat, the guesswork, or the platform dependency. Just run ag-refresh, ask away, and get answers that cite real files.

License

MIT License

Resources & Downloads

  • Source-code & Downloads

Related articles