Give your AI agent a memory that remembers your whole project and tells you what it does not know yet.
Before you start
This one is for the run-your-own-agents crowd. You should be comfortable running Claude Code from a terminal and have bun installed. No server, no cloud account, and no API-key wrangling to get started.
GBrain is built to be installed by an agent, not by hand. Open Claude Code in any project and paste the line below. Claude pulls GBrain's install protocol, runs it, and asks you for anything it needs as it goes. You answer questions, it does the work. About 30 minutes to a fully working brain, and the local database is ready in seconds.
Want to do it by hand instead? Spin up a local brain and connect it to Claude Code in two commands. The first builds a local database in about two seconds with no Docker and no server. The second registers GBrain as an MCP tool your agent can call. Zero server, zero token, zero tunnel. Swap claude for codex if you run Codex instead.
GBrain earns its keep on the stuff that lives outside your repo: meeting notes, decisions, the why behind a refactor, who owns what. Hand that context to your agent and let GBrain ingest it. As it writes each page it extracts the people, companies, and projects and links them with typed edges automatically, no extra LLM calls. That graph is what lets it answer questions plain search cannot.
This is the part that changes how you work. Ask a question that spans several notes. GBrain returns a synthesized answer with a source behind each claim, then ends with what it does not know yet. That closing note is the point: it tells you where to go look instead of letting your agent confidently fill the gap with a guess.
GBrain is a daemon, not a one-shot tool. It runs a dream cycle that ingests new material, enriches the people and companies it already knows, fixes its own citations, and consolidates memory while you are not looking. Set it up once and the brain gets sharper on its own. Everything stays on your hardware and your keys the whole time.
Once the MCP tool is wired in, these are the calls you reach for most. Your agent invokes them; you just ask in plain language.
| Command / Skill | What it does | When to use it |
|---|---|---|
gbrain thinkadvanced | Returns a synthesized, cited answer plus a note on what it does not know yet, not a page list. | Any question that spans more than one note or needs an actual answer. |
gbrain init --pglite | Builds a local brain in about two seconds with no server and no Docker. | First-time setup when you want everything on your own machine. |
gbrain serveadvanced | Runs the brain as an MCP server your coding agent connects to. | Wiring GBrain into Claude Code or Codex as a tool. |