Introduction to the MCP Gateway

The MCP Gateway sits between your AI apps and the MCP servers they call. Users connect their AI app to a Quilr gateway URL instead of the server itself, and every tool call is checked against your tool rules and guardrails, then logged with the user, the agent and the result.

Your AI app

Claude, ChatGPT, Cursor, VS Code

Connects to a Quilr gateway URL

QuilrAI MCP Gateway

Tool rules and guardrails

Access by agent, group and user

Logs, analytics, health

MCP servers

Remote servers from the MCP Library

Your own servers and APIs

Local packages on the user's computer

QuilrAI

Key concepts

Term	What it is
MCP server	The unit you configure. It holds its connection, sign-in, tools, guardrails and rules. Add one with Add MCP server or from the Library.
Quilr gateway URL	The address agents call for one server, for example https://mcpgateway.quilr.ai/github/mcp. The gateway calls the server's upstream URL.
OneMCP	One URL that serves every MCP a user may use. See OneMCP.
Allowed agent	An AI client, matched by its user-agent keyword. You choose which servers each agent may reach. See Agents Configuration.
User dashboard	Where users sign in at mcpgateway.quilr.ai, connect their accounts and copy their URLs.

New servers start with their tools turned off, so nothing is reachable until you enable it.

The MCP Gateway page

Go to Settings > AI Gateway > MCP Gateway.

Button	What it opens
Overall analytics	The MCP Gateway workspace for all servers, on the Analytics tab
OneMCP endpoint	OneMCP settings. See OneMCP.
Library	The MCP Library. A count shows pending install requests from users.
Allowed Agents	Which AI clients may connect, and which servers each one sees
Add MCP server	The Add MCP drawer. See Adding MCP Servers.
...	Connections (who has connected to which server) and ZIA integration (see Web Search Policy)

Summary tile	Shows
Tool calls	Calls in the period, the share that came through OneMCP, people calling, tools called, success rate
Estate	Servers that are Serving, Awaiting connection or have No tools, and how many are enabled or disabled
Tools reachable	Tools agents can call, with Scoped rules, OAuth servers and the Library queue
Stopped by guardrails	Calls the gateway refused, Guardrail flags, Failed calls, p95 latency

Click a row on a tile to filter the server list or open the matching view.

Banner	Action
N OAuth servers need an administrator connection before their tools can be discovered.	Review all filters the list to those servers. With one server, the button reads Connect and the server name.
N quick fixes could prevent failed tool calls	Review fixes opens the quick fixes drawer. See Input Aliases.

Below the banners, Elsewhere links to Inventory & analytics, the User dashboard and this Documentation. The server list has All, Local, Remote and API tabs, Search servers and a Filter for state, auth, source and runtime.

The server card

Each configured server has a card.

Area	What it shows
Header	Name, System, Library or Local package tag, slug and description
Endpoints	Quilr gateway URL (copy it for clients) and the Upstream URL, with upstream p95 and failures
Status	Serving, Awaiting connection, No tools found or Disabled, plus transport and auth
Chips	Tools, Guardrails, Scoped rules, Token saving. Each opens that section.
Actions	Show tools, Uninstall, and the enable toggle
Traffic	Calls over time, users, tokens saved, top tools
Inspect	Logs (the server's activity), Open in Inventory
Configure	General, API (API servers only), Guardrails, Token Saving, Tools, Group & User Rules

A server waiting for an administrator OAuth sign-in shows Connect OAuth on its card.

The workspace

Overall analytics and every Inspect and Configure button open the MCP Gateway workspace. Select one or more servers with the server picker, and change the period at the top.

Tab	What it shows
Analytics	Tool calls, Blocked, Guardrail flags, Via OneMCP, Tokens saved, Success, P95 latency, Users, Servers; calls over time and usage by server
Activity	Tool calls (one row per call), Interactions (calls grouped into conversations), Findings (guardrail detections)
Connections	Which users have connected to which servers, and their tool calls
Health	Gateway verdict, Rejected requests, Access denied, Auth challenges, OAuth refreshes, Upstream server errors, and rejection reasons. Covers the whole gateway.
Settings	The server configuration. Available for one server at a time.

Settings sections

Sections vary by server. Edits are drafts until you click Save settings.

Section	What you set
General	Name, upstream sign-in, custom headers, Access Control, Claims forwarding, Web Search Policy (QuilrAI Web Search only), Uninstall
API	Base URL, spec and access rules of an API MCP
Permissions	OAuth Permissions: the scopes the connection requests (OAuth servers with selectable scopes)
Tools	Tool visibility, Human approval, Tool Change Watch
Guardrails	Security Guardrails on tool arguments and tool results
Token saving	Token Saving strategies
Input aliases	Input Aliases: translate mismatched inputs from AI clients
Group & User Rules	Group & User Rules: overrides for a smart group or a single user
API tokens	API Tokens for direct connections (not on OAuth servers)

Policy Engine

When the Policy Engine is on, Settings shows Policy Engine active and marks Tools, Guardrails, Token saving and Group & User Rules with its icon. Those sections follow published policies instead of these settings. General and API tokens are still managed here.

Next steps

• Quick Start: add a server and make a first tool call.
• Adding MCP Servers: remote servers, sign-in modes and the MCP Library.
• API to MCP: turn a REST API into tools.
• OneMCP: one URL for every MCP a user may use.
• Integration Guide: URLs, tokens and client examples.