GBrain: The 73-Skill Second Brain for Claude Code

What is gbrain, in one line?

gbrain describes itself as the layer that turns raw material into an answer instead of a pile of pages. Flowy indexes the plugin from its public repository, garrytan/gbrain, which sits in the data category and carries 29,077 GitHub stars as of the last crawl. Flowy does not host or bundle the code itself, so the real install command lives at the listing page linked at the top of this guide. The plugin ships 73 components, all of them skills, with no agents, commands, MCP servers, or hooks. Together they turn Claude Code into a personal or company knowledge base the plugin itself calls "the brain." If you already keep notes in Obsidian, Notion, Logseq, or plain markdown, migrate is the skill that carries it over.

How does content get into the brain?

capture is the one command meant to replace ad hoc saving, and ingest looks at whatever you hand it and delegates: media-ingest for video, audio, PDF, books, screenshots, and GitHub repos; voice-note-ingest for a spoken note kept in the speaker's exact words; blog-ingest for a whole feed archive; idea-ingest for a link or a "read this" moment; meeting-ingestion for transcripts from any recorder; conversation-archive for AI chat exports. archive-crawler sweeps a personal file dump such as Dropbox or a Gmail takeout for anything worth keeping, and refuses to run without a gbrain.yml file telling it what counts. bulk-ingestion handles all of that at archive scale, cold-start sequences the first sources for a brand-new brain, and webhook-transforms turns an external event like an SMS into the same kind of signal.

What stops the brain from filling up with garbage?

brain-ingest-gate and brain-taxonomist run before a page is written, one deciding whether it belongs and the other where it goes. frontmatter-guard repairs malformed YAML, and citation-fixer keeps every fact traceable to a source. maintain runs the periodic health pass: back-link enforcement, stale-info detection, orphan pages, benchmarks. data-loss-gate sits in front of any bulk delete, and correction-pipeline treats a factual correction as a bug report and root-causes it rather than editing one page. context-audit checks always-loaded files such as CLAUDE.md and AGENTS.md for drift, and eiirp is the seven-phase pass that files away a finished work session.

How do you get an answer back out?

query answers from the brain with layered search, synthesis, and citations. retrieval-reflex is the habit of opening the relevant page before answering from memory, and resolve-before-asking exhausts the brain's own lookup chain before an identity question reaches you. idea-lineage traces one idea from its first mention to its current version. briefing compiles a daily readout from meeting context and active deals; daily-task-prep adds calendar lookahead and yesterday's open threads. daily-task-manager runs the add, complete, and defer cycle with stable IDs, cron-scheduler handles staggering and quiet hours, and reports stores timestamped output behind a link check before delivery.

What does the research layer cover?

academic-verify traces a claim through publication, methodology, raw data, and replication, handing the web lookup to perplexity-research, which reports back only what's new. strategic-reading reads a source through the lens of a specific problem and produces a playbook, not a summary. research-compendium archives every primary source, summarizes each one, then synthesizes one standalone page, and citation-graph-ingest builds a typed reference graph so the brain knows one document overrules or relies on another. fact-check checks every claim in a piece of content against live sources before it ships, and cross-modal-review spawns a second AI model to review it first. article-enrichment turns a raw article dump into a structured page, enrich keeps person and company pages current as new entities come up, and company-brainify turns a personal brain into a sanitized, shareable one.

Is there a voice layer?

voice-persona-venus is the default, built for fast logistics such as calendar and tasks over a live call, and voice-persona-mars takes over for reflective conversation or a demo. voice-post-call turns the finished call's transcript into a brain page afterward, and draft-in-voice ghostwrites short content in a specific person's validated voice.

How does gbrain maintain its own skills?

skillify, skill-creator, and skill-autobench build and evaluate skills from real usage rather than a specification. testing, smoke-test, and skillpack-check report on install and test-suite health, while gbrain-advisor and gbrain-upgrade flag drift and apply updates. schema-author and schema-unify manage the underlying page-type taxonomy, and setup handles the first install. minion-orchestrator steers shell jobs and subagent work, and two-tier-extraction splits a large corpus between a fast model and a slower one. publish and brain-pdf get a finished page back out as a password-protected HTML file or a PDF.

Who is this for?

Anyone running Claude Code as a long-lived personal or company assistant, where the value comes from what it remembers across sessions rather than any single conversation.