CogniLedger MCP Server — Makuri showcase

A public, read-only Model Context Protocol server operated by CogniLedger Solutions S.R.L. (Bucharest, Romania). It exposes structured metadata about the Makuri EdTech platform — 11 tools (9 info tools + 2 interactive MCP Apps panels), 3 markdown resources, and 2 guided prompts, covering mission, languages, teaching approach, pricing, safety, compliance posture, tech stack, contact channels, and free public learning resources.

This is a reference deployment demonstrating production MCP patterns under EU compliance constraints. Makuri is a High Risk AI system under EU AI Act Annex III, paragraph 3 (educational AI for minors); the v1 scope of this server is therefore deliberately narrow: metadata only, no user data, no PII, no aggregated analytics.

• Production endpoint: https://mcp.cogniledger.eu/mcp

• License: MIT

• Repository: github.com/Cogniledger/cogniledger-mcp-makuri

• Contact: leonid@cogniledger.eu

What this server is

• Public, unauthenticated, read-only

• 11 tools returning static metadata (no database queries against user data), including 2 interactive MCP Apps panels that render inline in supporting hosts

• 3 markdown resources (manifesto, child-safety overview, connect guide) and 2 guided prompts

• Designed to be called by AI assistants — Claude Desktop, Le Chat (Mistral), Cursor, ChatGPT Apps SDK, and any other MCP-capable client

Related MCP server: gitbook-mcp

What this server is not

• It is not the Makuri product. End users of Makuri (children, parents) interact with makuri.eu, not this MCP server.

• It does not expose user data, PII, IP addresses, behavioral analytics, or any data derived from end-user activity.

Connect

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "cogniledger-makuri": {
      "url": "https://mcp.cogniledger.eu/mcp"
    }
  }
}

Restart Claude Desktop. The tools appear under the connector picker.

Le Chat (Mistral)

In Le Chat settings, add a new MCP connector:

• Name: CogniLedger — Makuri

• Transport: streamable-http

• URL: https://mcp.cogniledger.eu/mcp

Cursor

Edit ~/.cursor/mcp.json:

{
  "mcpServers": {
    "cogniledger-makuri": {
      "url": "https://mcp.cogniledger.eu/mcp"
    }
  }
}

Tools

Resources

The server exposes three markdown documents via MCP resources/list / resources/read:

• makuri://docs/manifesto — the founder-written manifesto on why Makuri exists.

• makuri://docs/safety-overview — child-safety design measures (account model, data minimization, AI behavior controls).

• makuri://docs/connect-guide — how to connect this server in ChatGPT, Claude, and Le Chat.

(The two interactive panels above are also exposed as resources at ui://makuri/how-it-works and ui://makuri/romanian-quiz, for hosts that support MCP Apps.)

Prompts

Two guided prompts via prompts/list:

• evaluate_makuri_for_my_child(child_age, native_language) — guided fit evaluation for a specific child.

• makuri_safety_briefing() — honest safety briefing with explicit "design posture, not certified compliance" framing.

Full input schemas and example responses: docs/TOOLS.md. Real client transcripts: docs/EXAMPLES.md. Compliance disclosure: docs/COMPLIANCE_DISCLOSURE.md.

Local development

git clone https://github.com/Cogniledger/cogniledger-mcp-makuri.git
cd cogniledger-mcp-makuri
npm install
npm run dev

The server starts on http://localhost:3000. The MCP endpoint is http://localhost:3000/mcp.

Smoke test

# In one terminal:
npm run dev

# In another:
npm run test:smoke

The smoke test connects to the local server, lists tools, calls each one, and exits non-zero on any failure.

MCP Inspector

npx @modelcontextprotocol/inspector

Connect with transport streamable-http to http://localhost:3000/mcp. All 11 tools should be visible and callable, plus the 5 resources and 2 prompts.

Security and compliance

• No secrets in the repository. .env is git-ignored; .env.example lists variable names only and is empty for v1.

• No outbound API calls, no database queries, no authentication. v1 is pure static-data read.

• Per-tool structured logging emits one JSON line per invocation with five fields only: evt, tool, ts, status, duration_ms. No request bodies, no IPs, no argument values.

• See docs/COMPLIANCE_DISCLOSURE.md for the full compliance posture.

License

MIT — see LICENSE.

Copyright © 2026 CogniLedger Solutions S.R.L.