Documentation

Everything you need to get started with Kandela.

🎬 See Kandela in Action

☁️ Hosted (Cloud)

The fastest way to start. We handle the server — you just install the client.

Sign up for beta, get an API key, run one command. Takes 2 minutes.

Step-by-step guide →

What gets installed: MCP server registration, 8 lifecycle hooks (auto-recall, auto-save, prompt guard, etc.), and 16+ slash commands.
Compatibility: Currently supports Claude Code. Support for more AI coding tools is planned.

🏠 Self-Hosting

Run Kandela on your own server. Full control over your data.

Run Kandela on your own infrastructure with a few simple commands.

  • Requirements: Docker 20.10+, 4GB+ RAM
  • Setup: A few commands to pull, run, and connect
  • Your data stays on your server — no external dependencies

📦 Full guide & source code: github.com/deep-on/kandela-selfhost

🔧 Technical Overview

Built on multiple patent-pending technologies for memory retrieval, session continuity, and decision protection.

Hybrid Search

Semantic (ChromaDB) + BM25 keyword search, fused with Reciprocal Rank Fusion. Works across 50+ languages.

Importance Engine

25 rules auto-score memories 1–10. Infrastructure decisions get boosted; trivial facts decay over time (ACT-R-inspired exponential decay).

Prompt Guard

Detects when your agent suggests reversing a past decision. Surfaces the original reasoning before you regret it.

Circuit Breaker

After 3 repeated failures, stops execution and surfaces relevant gotchas from memory.

Lazy Retrieval

Brief mode (~260 tokens) at session start. Full context loaded on-demand via context_search.

Multi-Project

Isolated memory per project with cross-project discovery. Global memories shared across all projects.

A/B Benchmark Results

90 sessions, 2 scenarios (HIPAA pipeline + Slack infra bot), Claude Sonnet 4.6:

-79.5pp Decision Mistakes
-37% Code Bloat
-21% Time Wasted

🔌 Compatibility

Kandela works with any MCP-compatible client — Claude, GPT, Gemini, or any LLM. Switch between models freely; your memories stay the same. One memory store across all your tools.

Feature Claude Code
CLI / Desktop App		 Claude Desktop
Chat App		 Cursor / VS Code
& other MCP clients
13 MCP Tools (store, search, recall...)✅✅✅
Manual Operation (call tools directly)✅✅✅
Telegram Bot✅✅✅
Auto-Recall (session start)✅——
Auto-Save (session end)✅——
Prompt Guard✅——
Circuit Breaker✅——
Slash Commands (/kd-*)✅——
CLAUDE.md Guide✅——
One-click Install✅ManualManual

Claude Code CLI / Desktop App — Full experience. All features work automatically after one-line install.

Claude Desktop / Cursor / Other MCP clients — MCP tools + manual operation + Telegram bot. Set up manually below.

Setup: Claude Desktop

Add to ~/.claude.json:

{
  "mcpServers": {
    "kandela": {
      "type": "http",
      "url": "https://api.kandela.ai/mcp",
      "headers": {
        "Authorization": "Bearer mcp_YOUR_API_KEY"
      }
    }
  }
}

Setup: Cursor / VS Code (Claude Extension)

Add to .mcp.json in your project root. Restart the editor after saving.

{
  "mcpServers": {
    "kandela": {
      "type": "http",
      "url": "https://api.kandela.ai/mcp",
      "headers": {
        "Authorization": "Bearer mcp_YOUR_API_KEY"
      }
    }
  }
}

Setup: Other MCP Clients

Any MCP-compatible client can connect to Kandela. Use the HTTP transport:

Replace mcp_YOUR_API_KEY with your API key from Dashboard → Account.