LearnThat MCP

Learn while the agent codes.

LearnThat gives your coding agent a small learning loop.

Install<br>See why

One MCP URL. Your normal coding agent.

Add LearnThat as a remote HTTP MCP server. The agent discovers the tools,<br>opens OAuth when required, then weaves small learning moments into the work.

Codex<br>Claude Code<br>Cursor<br>Generic

Run in a terminal

Copy

codex mcp add learnthat --url https://learnthat-mcp-3bqygzrtsa-uc.a.run.app/mcp<br>codex mcp get learnthat

After login, restart or open a fresh Codex session so the tools are available.

Add at user scope

Copy

claude mcp add --scope user --transport http learnthat https://learnthat-mcp-3bqygzrtsa-uc.a.run.app/mcp<br># In Claude Code, run /mcp and choose Authenticate<br>User scope makes the server available across projects.

Remote MCP config

Copy

"mcpServers": {<br>"learnthat": {<br>"url": "https://learnthat-mcp-3bqygzrtsa-uc.a.run.app/mcp"<br>Use this shape for clients that read an mcpServers JSON config.

HTTP server entry

Copy

"servers": {<br>"learnthat": {<br>"type": "http",<br>"url": "https://learnthat-mcp-3bqygzrtsa-uc.a.run.app/mcp"<br>If your client asks for a server URL directly, use https://learnthat-mcp-3bqygzrtsa-uc.a.run.app/mcp.

Agents can make people faster, but weaker at reasoning through code.

Coding agents remove friction. That is useful, but it can also skip the<br>work where mental models form: predicting behavior, reading errors,<br>writing the one line that matters, and explaining why a change works.

For new builders

The app ships, but the codebase stays opaque when the agent owns every decision.

For engineers

Velocity rises, but unfamiliar systems can still turn into copy-and-accept work.

Small checks, placed where they matter.

LearnThat does not take over the coding agent. It gives the agent coaching<br>tools: assess level, plan checkpoints, generate micro-challenges, grade<br>short answers, and record useful snippets for later review.

Before the change

Ask for a prediction or risk call tied to the current task.

During the change

Invite the user to write or explain a tiny meaningful piece.

After the change

Record concepts, answers, snippets, and verification habits.

Keep the flow, add a little resistance.

A good prompt is short enough to answer without derailing the task. The<br>agent can ask for a prediction, a tiny snippet, or one sentence of reasoning,<br>then keep moving.

Soft prompts

Questions stay optional and tied to the file, error, or diff already in front of the user.

Useful memory

Answers and snippets become signals for the next challenge instead of dead chat history.

Soft by default. Different by level.

The agent keeps progress moving. Challenges adapt from the task, prompt<br>history, answers, skips, and quiz outcomes.

Measure retained concepts, not just completed tasks.

Events land in a durable analytics store so schools, teams, and solo builders<br>can see whether learning is improving over time.

Retention<br>Concepts answered again after a delay.

Quiz quality<br>Rolling score by topic, task type, and level.

AI dependency<br>Skipped or weak checks compared with strong completions.

Test LearnThat in your coding agent.

Copy the MCP URL, add it to your client, and let your next coding session include<br>small programming checks instead of passive code acceptance.

https://learnthat-mcp-3bqygzrtsa-uc.a.run.app/mcp

Copy MCP URL