Synlake — Cloud Infrastructure for AI Agents MCP server

Synlake MCP Server

The MCP connector for Synlake — B2AI infrastructure for autonomous agents. Synlake normalizes AWS, GCP, and Azure into one schema, generates ready-to-run execution kits (Terraform HCL + CLI), and prices them before deploy — so an AI agent can discover, evaluate, and provision cloud infrastructure with zero human intervention.

This repository is the open-source MCP server (a thin stdio ↔ HTTP proxy). The infrastructure intelligence — cross-cloud normalization, the cost engine, and execution-kit generation — runs as a hosted service at api.synlake.ai.

🌐 synlake.ai · 📖 Docs · 🔌 MCP: ai.synlake/synlake

An agent deploys infrastructure — no human in the loop

A coding agent needs a Kubernetes node pool. It doesn't open a cloud console, compare pricing pages, or hand-write Terraform. It asks Synlake, and gets back a deployable answer:

Agent → Synlake:  "compute, 4 vCPUs, 16 GB RAM, us-east-1, budget $100/mo"

Synlake → Agent:  best:  gcp · e2-standard-4 · $97.82/mo   (19.5% cheaper than the priciest)
                  also:  aws t3.xlarge $121.47 · azure Standard_B4ms $121.18
                  kit:   { terraform: "resource \"google_compute_instance\"…",
                           cli: "gcloud compute instances create…" }
                  ✓ validated: no public ingress, encryption on, within budget

The agent reviews the kit, applies the Terraform, and moves on. Every response is dense, deterministic JSON — built for machine consumption, not dashboards. Synlake tools are read-only: they recommend, price, and validate. A human (or the agent's own policy) runs the execution kit.

Architecture

flowchart LR
    A[AI Agent] -- MCP / REST --> B[Synlake API]
    B --> C[Normalizer]
    C --> D1[AWS adapter]
    C --> D2[Azure adapter]
    C --> D3[GCP adapter]
    C --> E[Ranking engine<br/>cost + constraints]
    E --> F[Execution-kit generator<br/>Terraform + CLI]
    F --> G[Guardrails & validation<br/>budget · region · security]
    G --> H[Audit log]
    H --> A

An agent sends an intent. Synlake normalizes options across clouds, ranks them by cost and constraints, generates a deploy-ready execution kit, runs guardrail checks, logs the call, and returns one machine-ready payload. This repo ships the open-source MCP connector (the AI Agent ↔ Synlake API edge); the boxes to the right of Synlake API run as the hosted service.

Quickstart

Hosted (recommended) — MCP over Streamable HTTP

Point any MCP client at the hosted server — three lines, nothing to install:

{
  "mcpServers": {
    "synlake": {
      "url": "https://api.synlake.ai/api/mcp",
      "headers": { "Authorization": "Bearer sk_synlake_YOUR_KEY" }
    }
  }
}

Local stdio wrapper (this package)

For clients that speak stdio (e.g. Claude Desktop), run the npm wrapper — it proxies to the same hosted server:

{
  "mcpServers": {
    "synlake": {
      "command": "npx",
      "args": ["-y", "@synlake-ai/mcp-server"],
      "env": { "SYNLAKE_API_KEY": "sk_synlake_YOUR_KEY" }
    }
  }
}

Get a free API key at synlake.ai (100 calls/month, no card required). The estimate endpoint is even callable without a key — zero friction to try.

Tools

Shortened. The full README is on GitHub.