tentra-mcp

Memory for AI coding agents. Persistent code graph + AI-generated architecture diagrams — MCP-native. Works in Cursor, Claude Code, Codex, and Windsurf.

Dogfood benchmark on our own monorepo: 99.4% token reduction (156.8× ratio) across 8 "where is X implemented?" queries — 114,644 tokens via file re-read vs 731 tokens via query_symbols. Full write-up →

Quick Start (60 seconds)

cd your-repo
npx tentra-mcp init --hook

One command:

1. Writes MCP config for Cursor / Claude Code / Codex / Windsurf (whichever are installed)
2. Installs a git post-commit hook so the code graph auto-refreshes after every commit — no manual re-indexing
3. Auto-derives your repo_id from the git remote and saves it to .tentra/metadata.json

Then grab your API key at trytentra.com/settings, replace YOUR_TENTRA_API_KEY in the generated config, reload your IDE, and ask your agent:

Index this codebase with Tentra and list the god-nodes

From here on, every git commit fires a background re-index. Your agents stay caught up automatically.

Skip the hook: drop --hook — just writes IDE configs. Manual stdio install: npx tentra-mcp (opens browser for GitHub device-flow auth on first tool call). API key instead: npx tentra-mcp --key YOUR_API_KEY.

What is Tentra?

Tentra is the persistent memory layer for AI coding agents. Describe a system — get a diagram and 14-framework code exports. Index your repo — agents query a structured graph of files, symbols, imports, and call edges instead of re-grepping source every session.

This MCP server gives your AI assistant 32 tools:

Architecture (9 tools)

Code Graph — Write (4 tools)

Code Graph — Read (10 tools)

Enrichment — Contracts, Decisions, Ownership, Domains (9 tools)

Setup

Option 1: SSE (zero install)

Add to your IDE's MCP config — no local install needed:

Cursor (Settings > Features > MCP > Add Server):

{
  "tentra": {
    "type": "sse",
    "url": "https://trytentra.com/api/mcp?key=YOUR_API_KEY"
  }
}

Claude Code (.mcp.json in project root):

{
  "mcpServers": {
    "tentra": {
      "type": "sse",
      "url": "https://trytentra.com/api/mcp?key=YOUR_API_KEY"
    }
  }
}

Option 2: Local install (needed for codebase scanning)

npx tentra-mcp

Authenticates automatically via GitHub on first use. Credentials are saved to ~/.tentra/credentials.

Cursor config for local server:

{
  "tentra": {
    "command": "npx",
    "args": ["tentra-mcp"]
  }
}

Claude Code (.mcp.json):

{
  "mcpServers": {
    "tentra": {
      "command": "npx",
      "args": ["tentra-mcp"]
    }
  }
}

Usage Examples

Once connected, just talk to your AI:

"Design a payment system with Stripe, Kafka, and PostgreSQL"
→ AI calls create_architecture → diagram at trytentra.com/arch/xxx

"Scan this codebase and generate the architecture"
→ AI calls analyze_codebase → detects services, DBs, queues

"Export this architecture to Java Spring Boot"
→ AI calls export_architecture → downloads zip with project scaffold

"What changed since last time? Is my diagram outdated?"
→ AI calls sync_architecture → drift report with accuracy score

Export Formats

Java (Spring Boot), Node.js (Fastify), Python (FastAPI), Go (chi), Rust (Axum), .NET (ASP.NET), Kotlin (Ktor), PHP (Laravel), Ruby (Rails), Elixir (Phoenix), Docker Compose, Mermaid, ADR, Terraform

Links

• Website: trytentra.com
• Documentation: trytentra.com/docs
• Setup Guide: trytentra.com/docs/setup
• Gallery: trytentra.com/gallery

Development

This repo contains the open-source MCP server. The Tentra API and web app are a separate hosted service at trytentra.com.

npm install --legacy-peer-deps
npm run build      # tsc --noEmit + esbuild bundle → dist/index.js
npm start          # run the bundled server
npm test           # vitest

The published npm package (tentra-mcp) ships only the bundled dist/ — source is here for auditability and community contributions.

License

MIT