What is an MCP gateway? MCP proxies compared

What does an MCP gateway do?

Without a gateway, every agent connects to every server on its own. Each connection has its own sign-in, credentials and logs. A gateway puts one endpoint in front of all of them.

• Aggregation: many servers appear as one, with one tool list.
• Sign-in: the agent authenticates once to the gateway.
• Credential injection: the gateway holds upstream keys or tokens, so the agent never sees them.
• Tool control: admins hide, rename or allow tools per user or team.
• Rate limits and quotas per user, team or tool.
• Logs and traces of every tool call.
• Translation: some gateways turn REST APIs or Lambda functions into MCP tools.

How do the main MCP gateways compare?

They split into two groups. Self-hosted gateways run on your infrastructure. Managed gateways are run by a vendor. Facts come from each vendor page on November 10, 2026.

What must an MCP proxy check?

The MCP spec has rules for proxies because they hold other people’s tokens. Check any gateway you buy or build against this list.

• No token passthrough. A server must not accept a token issued for another service, or forward the client’s token upstream. The gateway uses a separate token issued for the upstream server.
• Per-client consent. A proxy that uses a static client ID with a third-party API must keep a record of which clients each user approved. Otherwise an attacker can reuse a consent cookie to steal an authorization code.
• Exact redirect URIs. Match them as exact strings, with no wildcards.
• Header and body must agree. Under 2026-07-28, Streamable HTTP copies the method and tool name into Mcp-Method and Mcp-Name headers so a gateway can route without parsing JSON. Servers must reject a request whose headers and body differ.
• Name collisions. Two servers can both expose a tool called search. The tools spec tells aggregators to tell them apart, for example with a server ID prefix.

Is an MCP gateway the same as an LLM gateway?

No. An LLM gateway sits between your app and model providers such as OpenAI or Anthropic. It routes prompts, tracks tokens and fails over between models. An MCP gateway sits between the agent and its tools.

Some products do both. agentgateway routes LLM, MCP and agent-to-agent traffic in one binary. AgentCore Gateway also routes inference across model providers.

Locus Pro is a metered tool gateway, not an LLM gateway. It does not sit between your agent and its own model provider. It sits between your agent and the paid APIs the agent calls.

What does an MCP gateway cost?

Self-hosted gateways are free software, but you run and secure the servers. Managed gateways charge for traffic or seats. With most of them, the tools behind the gateway still bill on their own.

AWS AgentCore Gateway bills per use: $0.005 per 1,000 tool calls, $0.025 per 1,000 tool searches and $0.02 per 100 tools indexed a month. The APIs you connect behind it bill separately. Prices checked November 10, 2026.

Locus Pro works differently, because the paid APIs sit inside it. Each call is charged its catalog price in credits from one prepaid balance, with no provider accounts to pay separately. The Free plan has no monthly fee and allows up to 5 paid tool calls a minute. Paid plans add included credits and higher limits: 15 calls a minute on Plus and 240 on Enterprise.

How does Locus Pro work as a gateway for paid APIs?

Locus Pro is one hosted MCP endpoint, https://api.paywithlocus.com/api/credits/mcp, in front of many paid APIs. Locus holds the provider accounts. Your agent needs no provider keys.

The compact surface lists four tools: search, describe, execute and get_result. The agent searches the catalog, reads one tool’s schema and price, then executes it. Search and describe are free.

• Sign-in: OAuth per user, with mcp:read for discovery and mcp:execute for paid calls.
• Price ceiling: each execute call can carry max_credits. A call priced above it fails and nothing is charged.
• Spend caps: opt-in. Set a workspace monthly hard limit, a monthly budget per agent key or connected app, or approval for calls over a set amount.
• Call history: list_calls returns each call with its status and credits charged.
• Per-customer billing on Enterprise: bill each of your customers from their own balance at base price plus your markup.
• Limits: it does not aggregate MCP servers you already run. Enterprise workspaces can add their own HTTP APIs as Custom APIs.

• See the Locus Pro tool catalog
• Locus Pro plans and rate limits

Which MCP gateway should I choose?

Start from what sits behind the gateway and who runs it.

• Local servers on developer laptops: Docker MCP Gateway, for container isolation.
• Internal MCP servers behind company SSO: Cloudflare MCP server portals or Kong AI Gateway.
• Your own APIs and Lambda functions on AWS: AgentCore Gateway.
• One open source data plane for MCP and LLM traffic: agentgateway.
• Paid third-party APIs with one bill and no provider accounts: Locus Pro.
• Tools inside your users’ own SaaS accounts: an OAuth connector platform such as Composio or Arcade.

• Composio alternatives
• Is there an OpenRouter for APIs?