Saylor (@Kev96790724) on X

What it claims

G-Krusch is a self-hosted bridge that lets coding agents and the author's Krusch Companion app reach Google Drive, Calendar and Gmail. OAuth tokens stay on the user's machine instead of passing through a SaaS middleman. Beyond the RAG and MCP server in the existing summary, the announcement describes one more feature: an agent.md steering file kept in Google Drive. A user can edit agent rules from a phone, and every IDE agent connected to the hub picks up the change automatically. Freshness uses what the author calls "mutation-epoch cache coherency."

How it works

PDF ingestion has three tiers behind a gate. Before spooling a file, the hub checks for the %PDF- magic bytes. Files that pass go first to the "KruschNexus" layout engine, then to Poppler, and finally to a pure-JS stream parser. Each tier emits page citations in the form [doc.pdf p.1]. The text then goes to the embedding cascade: local Ollama bge-large at 1,024 dimensions, then Gemini at 1,024 dimensions, then an "uncached offline fallback." This sits alongside SQLite FTS5 BM25 keyword search.

flowchart TD
  F[Drive file] --> G{"%PDF- magic bytes?"}
  G -- no --> X[Rejected before spool]
  G -- yes --> K[KruschNexus layout]
  K -- fails --> P[Poppler]
  P -- fails --> J[Pure-JS stream parser]
  K --> C[Text + page citations]
  P --> C
  J --> C
  C --> E{Embedding}
  E --> O[Ollama bge-large]
  O -- unavailable --> GM[Gemini]
  GM -- unavailable --> OF[Offline fallback]
  C --> B[FTS5 BM25]

The Gmail tools handle RFC 2822 messages and RFC 2047 encoded headers. The security measures listed beyond the token vault are:

• a CSPRNG-generated OAuth2 CSRF state pool;
• isFileInWorkspace ancestry checks that keep file access inside a workspace boundary;
• FIFO keyed mutexes that serialize concurrent writes.

The vault uses AES-256-GCM with 96-bit IVs, following NIST guidance.

How to use it

The post links an MIT-licensed GitHub repository under github.com/kruschdev/ and an architecture article on krusch.dev. Both URLs are truncated in the post text. The post gives no install commands, configuration steps, or Google Cloud OAuth client setup.

Limits

The source is a single launch post on X by @Kev96790724, with no date in the fetched text. The post gives no retrieval-quality numbers and no latency figures. It does not say how dense and BM25 scores are combined. It does not describe the "uncached offline fallback" embedding or say what KruschNexus is, and it does not say whether that engine is bundled. The only test evidence is "30/30 tests passing," with no stated coverage.