Concepts
An open-source MCP for your coding agent

Learn from the code
your AI just wrote.

After every commit, your agent draws the change as a system-design diagram and marks each concept it relies on, right where it lives.

Add the MCP to your agent
claude mcp add --scope user concepts -- npx -y concepts-mcp@latest

Needs Node.js 20+ and git. Paste into your terminal.

Then turn it on in a repository. Ask your agent, or run the setup yourself; it asks where diagrams go and how much each commit saves:

“Set up concept diagrams for this repo.” or npx -y concepts-mcp@latest init

Free and open source · Runs on your machine · Needs Node.js 20+ and git · No account or API key

3f9c2ab · Retrying failed webhooks · concepts.svg
A hand-drawn diagram titled Retrying failed webhooks. An architecture view shows a customer calling the billing API, which writes to a new webhook outbox in the same transaction; a new delivery worker polls the outbox, posts to the customer endpoint, signs with stored keys and gives up into dead letters. A sequence view shows one failed attempt. Six numbered concept markers sit on the parts where each concept lives.
  1. 1Transactional outbox
  2. 2Exponential backoff
  3. 3Jitter
  4. 4Idempotency keys
  5. 5Dead-letter queue
  6. 6HMAC signatures

From commit to understanding.

The MCP reads git and draws the files. The agent you already use picks out the concepts.

01

Connect the MCP

One command adds Concepts to your coding agent. It runs locally over stdio.

02

Your agent models each commit

It reads the diff, chooses the views that teach it, and lists every concept worth learning.

03

You learn what it used

A diagram, an explanation per concept, and a note that grows each time it shows up again.

Diagrams, not paragraphs.

Five views, as many as the change needs, with real shapes for databases, queues, people and decisions. Explanations go beside the diagram, never in it.

Architecture

Components and how they talk

Sequence

Who calls whom, in order

Flow

Steps and decisions

State

States and transitions

Data model

Entities and relationships

A library that grows
with every commit.

Each concept gets one Markdown note that collects every commit where it showed up. Notes you add yourself are kept. You choose where it lives during setup.

  • projectdocs/concepts/shows up on GitHub
  • library~/Desktop/ConceptLibrary/just for you
Browse this repo’s library →

What your agent gets.

Your agent does the thinking. The MCP gives it the workflow, the history and a renderer.

get_concepts_instructions

The workflow, so every commit is modeled the same way.

setup_concepts

Sets up or changes where diagrams go and how much each commit saves, with a picker.

get_unprocessed_commits

Commits without a diagram yet, with diffs sized to fit.

get_concept_library

Concepts already learned, so their notes are reused.

record_commit_concepts

Lays out and draws the diagrams, then updates the library.

Private by design

Runs on your machine. Your code never leaves it.

Make every commit a lesson.

One command, then keep coding.