Skip to content
How it works

CKB works behind your agent

You talk to your coding agent the way you always do. CKB sits underneath it, feeding it evidence about your codebase.

You
Your coding agent
CKB
Your codebase

In the normal coding workflow, CKB works behind your agent. It does not replace it.

We call the tool the developer already uses — Claude Code, Cursor, Windsurf, VS Code and other supported agents — a consumer. A consumer talks to CKB the same way it talks to any other MCP server: it asks a question, CKB answers with evidence gathered from your codebase. You never talk to CKB directly.

What happens at setup

Four commands to get started. Keep the index fresh afterward with `ckb mcp --watch` or the daemon (`ckb daemon start`).

1npm install -g @tastehub/ckb

Install the CKB CLI.

2ckb init

Set up CKB in your project.

3ckb index

Build the symbol index CKB queries from.

4ckb setup

Connect CKB to your coding agent.

How your agent picks capabilities

CKB exposes 110 capabilities in total. Your consumer doesn't load all of them — it starts with a default set of 25 core capabilities.

When a task needs something outside that default set, your agent can expand its access to the full 110. CKB's next release tells your agent when to do that on its own.

Evidence sources

CKB layers multiple backends. Not all of them are available in every environment — here's what each one needs.

SourcePreconditionWhat it gives you
GitAlways available.History, blame, co-change — works the moment your project is a git repo.
SCIPNeeds ckb index and a language indexer.A compiler-grade symbol index for precise navigation and references.
LSPNeeds a language server on PATH.Falls back to whatever language servers you already have installed.
CartographerBuilt into macOS and Linux release binaries. Not available on Windows.Structural analysis — module boundaries, architecture, coupling.
LIPOptional, separate install.Live introspection for languages where static analysis alone falls short.

CKB never modifies your source code

CKB analyzes and explains your code — your agent decides what to change, CKB only tells it what it's looking at. A few capabilities do write, but only into CKB's own store, never your codebase: recording a decision (recordDecision) or annotating a module writes a Markdown file and database entry inside CKB's data, not your source tree.

See it in practice

Real prompts and what CKB contributes to each one.