Connect agents to any tool

HOW IT WORKS

Everything you need to manage MCP servers

Six Server Types

Support for Remote (HTTP/SSE), NPX (Node.js), UVX (Python), Container (Docker/K8s), Generated (AI-created from API docs), and Bundle (curated collections). Connect to any MCP server regardless of how it's hosted.

Credentials That Never Expire

The gateway keeps OAuth tokens alive indefinitely. A background service refreshes tokens before they expire — with intelligent retry, exponential backoff, and rate-limit awareness. Your agents authenticate once and stay connected forever. No human re-authentication, no stale tokens, no surprise auth failures.

BACKGROUNDtoken-refresh-daemon

▸ Checking tokens due for refresh...

▸ Found 3 tokens expiring within 5 min

✓ "github" refreshed · next in 55m

✓ "office365" refreshed · next in 58m

✓ "slack" rotated · next in 52m

All 47 connections healthy · 0 failures

AES-256-GCM encrypted · per-user isolation · runs every 60s

Deep Server Management

Drill into any server to see its tools, OAuth connections, sync status, and credential health. Enable or disable individual tools, search across tool schemas, and sync on-demand to pick up upstream changes.

Search, Then Execute

The biggest problem in production AI: tool sprawl. Anthropic says tool accuracy degrades past 30 tools. Cursor hard-caps at 40. With 5 MCP servers averaging 30 tools each, you burn 55K+ tokens on tool definitions before the agent does any work. MCP Gateway solves this with progressive tool loading. Instead of dumping every tool into the context window, the agent gets two meta-tools: SEARCH_TOOLS and EXECUTE_TOOL. Search by intent, get back the top matches with name and schema, then execute the one you need. Thousands of tools, two context slots. Configurable top-K results, three gateway modes (LIST, SEARCH+EXECUTE, AUTO), and the threshold is a setting you control.

POST/mcp/gateway

// Step 1: Agent searches by intent

{

"method": "tools/call",

"params": {

"name": "SEARCH_TOOLS",

"arguments": { "query": "send an email" }

}

}

200 OK18ms

// Step 2: Agent picks the best match and executes

[

{ "name": "O365__send_email",

"score": 0.96, "server": "office365" },

{ "name": "GMAIL__send_message",

"score": 0.89, "server": "google" }

]

247 tools across 12 servers → 2 context slots

AI Server Generation

Paste an API docs URL or OpenAPI spec. The Deep Agent analyzes the API, generates tool definitions, and deploys a working virtual MCP server. Review and approve before it goes live.

POST/api/v1/servers/generate

// Request

{

"input_type": "docs_url",

"input_content": "https://docs.stripe.com/api",

"server_name": "stripe-api"

}

// SSE stream

data: { "phase": "analyzing" }

data: { "phase": "enriching" }

data: { "phase": "complete", "tool_count": 18 }

One URL, Every Tool

Point any MCP client — Claude Desktop, Cursor, VS Code, Windsurf — at the unified gateway endpoint. One URL and one API key gives your agent access to every registered server.

GETmcp.json

{

"servers": {

"mcpgateway": {

"url": "https://your-domain.com/mcp/gateway",

"headers": {

"Authorization": "Bearer mgw_usr_..."

}

}

}

}

CONTEXT ENGINEERING

The right tool, at the right time

Anthropic's research shows tool accuracy collapses past 30-50 tools. Speakeasy benchmarks show static tool loading becomes infeasible at 200+ tools. With SEARCH+EXECUTE mode, MCP Gateway replaces hundreds of tool definitions with two meta-tools — bringing only what the agent needs into context, on demand.

Based on Anthropic, Speakeasy, and MCPVerse benchmarks (2025-2026)

API REFERENCE

Complete REST API

Every operation is an API call. Build integrations, automate workflows, or connect your own tools.