🔌 Agentic KDD MCP Server

CoALA Memory + Self-Healing — natively in Cursor and Claude Code

No more manual commands. No more skipped steps. Memory works automatically.

Install · Tools · How it works · Agentic KDD

Without MCP:  Claude "decides" to run terminal commands — may skip steps
With MCP:     Tools called automatically as native capabilities — always runs

The problem this solves

Agentic KDD has a full CoALA memory system — episodic, semantic, procedural layers — and a self-healing test loop. But without MCP, Claude has to "decide" to run the right terminal commands. In long sessions, it forgets. Steps get skipped.

With the MCP server, every tool is a native capability — like web search or file reading. Claude calls them automatically. No decisions. No skipping.

Install

# 1. Install Agentic KDD (if not already)
npm install -g agentic-kdd

# 2. Install the MCP server
npm install -g agentic-kdd-mcp

# 3. Register in Cursor + Claude Desktop + Claude Code
akdd-mcp install

# 4. Restart Cursor / Claude Desktop
# Done.

What changes

Before MCP — per project:

akdd init                    ← manual
akdd sync                    ← manual
aa: configurar               ← manual
# Claude may skip memory lookup
# Claude may skip test execution
# Claude may skip episodic registration

With MCP — once ever:

npm install -g agentic-kdd-mcp
akdd-mcp install
# Restart Cursor

Then in any project:

Open in Cursor → type aa: configurar
Everything else is automatic

Tools

8 tools available natively in Cursor, Claude Code, and Claude Desktop:

| Tool | What it does | Runs automatically | |------|-------------|-------------------| | kdd_status | Installation check + CoALA memory stats | On first aa: in a session | | kdd_buscar | Hybrid search: procedural + episodic + semantic | Before every implementation | | kdd_impacto | What breaks if you modify module X | Before touching any file | | kdd_sync | Sync memory → SQLite + decay + episodic consolidation | After every aa: cycle | | kdd_analizar | Analyze code → fill entities + semantic relations | On new projects | | kdd_ejecutar_tests | Execute real tests + return actual output | Self-healing loop | | kdd_registrar_episodio | Save episodic memory (what was tried, why it worked) | After every fix | | kdd_init | Initialize Agentic KDD in new project | First aa: in new project |

How it works

You open any project in Cursor
Type: aa: configurar

MCP calls automatically:
  kdd_status()      → is KDD installed? if not → kdd_init()
  kdd_analizar()    → reads code → fills semantic graph (entities + relations)
  kdd_sync()        → syncs memory to SQLite

Type: aa: fix the QR connection bug

MCP calls automatically:
  kdd_buscar("QR bug", "whatsapp")
    → searches procedural memory (known patterns + errors)
    → searches episodic memory (was this tried before? what worked?)
    → searches semantic memory (what depends on QR manager?)
    → returns top 10 results with relevance score in <5ms

  kdd_impacto("BaileysSession")
    → BaileysSession ← depends on ← SessionManager
    → touches BaileysSession → may break SessionManager
    → agent knows this before writing a single line

  [implementation happens with full context]

  kdd_ejecutar_tests()
    → runs npm test / npx jest / php artisan test
    → returns real output: N passed, N failed
    → if failures → self-healing loop triggers

  kdd_registrar_episodio({
    descripcion: "QR fix for Baileys 515",
    accion_tomada: "return early on 515 status code",
    resultado: "resuelto",
    razon_resultado: "515 is normal post-QR pairing, not an error"
  })
    → saved to episodic memory
    → next time this error appears → goes straight to solution

  kdd_sync()
    → episodic consolidation: resolved episodes → patterns
    → temporal decay: stale patterns lose relevance
    → everything in SQLite, ready for next cycle

CLI commands

akdd-mcp install    # register in Cursor + Claude Desktop + Claude Code
akdd-mcp uninstall  # remove from all IDEs
akdd-mcp status     # check where it's installed
akdd-mcp start      # start server manually (for testing/debugging)

Manual configuration

If akdd-mcp install doesn't cover your IDE, add manually:

Cursor (~/.cursor/mcp.json or via Settings → MCP)

{
  "mcpServers": {
    "agentic-kdd": {
      "command": "node",
      "args": ["/path/to/global/node_modules/agentic-kdd-mcp/server.js"],
      "description": "Agentic KDD — CoALA memory + self-healing"
    }
  }
}

Claude Code (terminal)

claude mcp add agentic-kdd -- node $(npm root -g)/agentic-kdd-mcp/server.js

Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "agentic-kdd": {
      "command": "node",
      "args": ["/path/to/global/node_modules/agentic-kdd-mcp/server.js"]
    }
  }
}

Requirements

• Node.js 18+
• Cursor, Claude Code, or Claude Desktop
• Agentic KDD: npm install -g agentic-kdd

Architecture

Cursor / Claude Code / Claude Desktop
         ↓ MCP Protocol (stdio)
  agentic-kdd-mcp/server.js
         ↓ calls
  .agentic/grafo/grafo.cjs   ← CoALA SQLite engine
         ↓ reads/writes
  .agentic/memoria.db        ← procedural + episodic + semantic + working memory

The server auto-detects the current project by walking up the directory tree looking for .agentic/. Works with any project structure.

Relation to Agentic KDD

This is the MCP layer for Agentic KDD.

agentic-kdd      → the framework (agents, memory, pipeline)
agentic-kdd-mcp  → connects the framework to your IDE natively

Both packages work independently. The MCP server enhances Agentic KDD by making every memory operation automatic — but Agentic KDD works fine without it.

Made with 🧠 by @Adrianlpz211

agentic-kdd · agentic-kdd-mcp · GitHub