A ChatGPT for your repo that actually cites line numbers

What it does

Antigravity is a cross-IDE knowledge engine that indexes your codebase into a local .antigravity/ folder. A cluster of module-specific agents reads every part of your project and writes structured knowledge docs; when you ask a question via ag-ask, a router dispatches it to the agent that actually owns that code. Answers come back grounded in real source files, complete with paths and line numbers, rather than repo-wide grep dumps.

The interesting bit

The portability angle is the real architecture: it is not a plugin but a file-based knowledge layer. Because the intelligence lives in plain Markdown under .antigravity/, any IDE—Claude Code, Cursor, Codex CLI, Windsurf—can read the same context without vendor lock-in. It is essentially treating your repository as a static site and every module as its own embedded librarian.

Key highlights

• Benchmarked (self-reported, n=36) at 99% factual accuracy on Python codebases, with 2.1× lower latency than Codex CLI on lookup questions.
• Each module gets its own agent during ag-refresh, producing a distributed knowledge graph rather than one monolithic context blob.
• Ships as slash commands inside Claude Code and Codex CLI, plus a standalone CLI and MCP server.
• Supports any OpenAI-compatible endpoint, including local Ollama.
• Offers a “context files only” mode that bootstraps IDE entry files without requiring an LLM.

Caveats

• Synthesis and high-level architecture questions scored 81% in the project’s own benchmark, well behind Codex CLI (100%) and Claude Code (94%).
• Requires an upfront ag-refresh build step that can take minutes; stale knowledge bases produce stale answers.
• Codex CLI support currently needs a manual engine install because Codex hooks are not yet supported.

Verdict

Worth a look if you juggle multiple AI IDEs and want a single, portable knowledge layer that beats raw agents on factual lookups. Skip it if your primary need is high-level architectural synthesis, where native Codex or Claude still outperform.