kbrain

Agent Architecture

Boost your agents with specialized, shareable knowledge bases

Learn why connecting AI agents to Google Drive is not enough. KBrain adds an intent layer and AI-targeted indexing that makes agents faster, smarter, and shareable at scale.

Build your first knowledge brain

Create a brain

You have built a Claude agent. You connected it to Google Drive and GitHub. It works. So you wonder: what does KBrain add on top of that?

The short answer is this: Google Drive gives the agent access to your files. KBrain gives the agent the intelligence to know which files matter, before it even reads them.

A real test: KBrain vs. Google Drive MCP

A developer set up a Claude agent accessible via Telegram, with MCP connections to both Google Drive and GitHub. They created a dedicated folder in Drive and connected it to KBrain to run a direct comparison.

First finding: in that initial test, retrieval quality looked similar between the two approaches. The agent found the right documents either way.

The reason the results looked similar: the brain descriptions were left empty. Without descriptions, KBrain has no pre-matched context to work from. It falls back to real-time search, which is functionally the same as Drive MCP. The intent layer only activates when brains are described.

What Google Drive MCP actually does

Google Drive MCP is a read/write connector to your personal Drive. It is excellent for what it does: letting an agent read files, create documents, and update spreadsheets on your behalf.

  • Personal scope: it connects to your Drive, not a shared knowledge layer
  • Read and write: the agent can modify files, not just retrieve them
  • Real-time search: every query triggers a live Drive search at request time
  • No intent matching: the agent must figure out which files are relevant on each request
  • Not shareable: you cannot publish a Drive connection for other users or agents to consume

For personal productivity, that is often enough. For agent systems that need to serve multiple users, or that need to respond reliably and fast, it is not.

The architecture without KBrain

Architecture — 01
Agent connected directly to Drive
Every request triggers a live search. The agent discovers what exists at runtime.
⚠️ Without KBrain
1
User sends message to agent via Telegram or API
2
Agent must decide which files in Drive are relevant — no guidance, no intent map
3
Agent performs a live Drive search: slow, scope-limited, dependent on Drive's search index
4
Agent reads candidate files and infers which information applies
5
Agent returns an answer — quality depends on whether it found the right files
6
Setup is personal: another user or agent cannot reuse this connection
?
no pre-matched context
discovery happens at runtime
The agent is capable. But it must rediscover the relevant knowledge on every request. The more files exist in Drive, the harder this gets.

What KBrain adds: the intent layer

KBrain sits between your data sources and your AI agent. It is not just a connector. It is a structured knowledge layer with three properties that Drive MCP does not have.

  • Brain descriptions: each brain has a description that tells the agent what it contains and when to use it — without reading a single file
  • Shareable by design: you can publish a brain to the KBrain marketplace or keep it private and share it with specific users or agents
  • AI-targeted indexing: KBrain pre-matches your data to intents so the right files are surfaced directly, without a live search

The brain description is the most important field you can fill in. It is what lets the agent route requests correctly before any retrieval happens. An empty description means the agent gets no pre-matched context.

The architecture with KBrain

Architecture — 02
Agent connected via KBrain
Intent matching happens before retrieval. The agent knows where to look before it looks.
⚡ With KBrain
1
User sends message to agent via Telegram, API, or any MCP client
2
Agent reads brain descriptions — knows which brain covers this topic before searching
3
KBrain routes to the right files — pre-indexed and intent-matched, not discovered at runtime
4
Agent gets a grounded, traceable answer — sourced from verified, curated content
5
Same brain works for any user — publish once, share across Claude, ChatGPT, any MCP client
✓
intent matched before retrieval
works across users and models
The brain description does the routing. When it is filled in correctly, the agent spends no time figuring out where to look. It goes directly to the right knowledge.

How the intent layer compares to Drive search

Comparison — 03
Google Drive MCP vs. KBrain
Two different tools solving two different problems.
📁Google Drive MCP
— Personal read/write access to your Drive
— Live search at request time
— No intent descriptions or pre-matching
— Not shareable with other users or agents
— Discovery happens at runtime, every time
1
user scope
personal only
🧠KBrain
→ Structured knowledge layer over any data source
→ Intent-matched before retrieval
→ Brain descriptions route requests without live search
→ Shareable: marketplace, private, or team scoped
→ AI-targeted indexing pre-matches data to intents
∞
user scope
any MCP client
They are complementary. Drive MCP is the read/write layer for personal workflows. KBrain is the knowledge layer for agents that need to serve multiple users reliably.

The indexing feature: what is coming

The intent layer today requires you to write brain descriptions manually. This is already a significant advantage over raw Drive search, but it is a manual step.

KBrain is building AI-targeted indexing: a process that pre-matches your data to likely intents automatically, so each agent request goes directly to the right files and tokens without any real-time search.

  • Currently: manual brain descriptions, tested and validated in early setups
  • Next: AI-generated descriptions triggered when data changes in connected sources
  • End state: zero-configuration intent matching — add a data source, the indexing happens automatically

Google Drive does not do AI-targeted indexing. Its search index is optimized for human keyword queries. KBrain indexes for agent intent — a fundamentally different optimization target.

When to use KBrain on top of Drive MCP

  • You want to share knowledge with other users or agents without giving them Drive access
  • You have more than a handful of documents and runtime search quality is unreliable
  • You want the agent to know where to look before it searches, not after
  • You are building for multiple users and need one knowledge source that works for all of them
  • You want the same brain to work on Claude, ChatGPT, and future MCP clients without reconfiguring

Drive MCP and KBrain are not competing tools. Drive MCP handles the read/write personal workflow. KBrain handles the knowledge layer for agents that need to be reliable, fast, and shared.

Build your first knowledge brain

Subscribe to KBrain, create a brain from your expertise or your data, and make it available to Claude, ChatGPT, or any MCP compatible assistant.

Create a brain

Frequently asked questions

Why did retrieval look the same in the first test?

The brain descriptions were left empty. Without descriptions, KBrain has no intent context to match against and falls back to real-time search, which behaves like Drive MCP. Filling in brain descriptions is what activates the intent layer.

Can I use KBrain with Google Drive as a data source?

Yes. KBrain can connect to Google Drive and index the files into a brain. The agent then queries KBrain rather than Drive directly, getting intent-matched retrieval instead of live search.

What is the intent layer?

Each KBrain brain has a description that tells the agent what it contains and when to use it. The agent reads these descriptions to route requests to the right brain before any retrieval happens. This is the intent layer.

Can multiple users share the same KBrain brain?

Yes. You can publish a brain to the KBrain marketplace or share it with specific users. Multiple agents and users can query the same brain simultaneously. This is not possible with a personal Drive MCP connection.

What is AI-targeted indexing?

AI-targeted indexing pre-matches your data to likely agent intents. Instead of searching at request time, the agent routes directly to the right files and tokens. This feature is in development and will eventually be triggered automatically when data changes.