Learn how to integrate an MCP gateway with Claude Code to consolidate tool access, enforce governance policies, and reduce token consumption across connected MCP servers.
Claude Code has emerged as a standard terminal-based coding agent for engineering teams. Its built-in support for the Model Context Protocol (MCP) enables interaction with filesystems, databases, GitHub, web search, Slack, internal APIs, and an expanding ecosystem of community-hosted tool servers. While connecting Claude Code to a small number of MCP servers is straightforward, scaling to dozens of servers introduces operational complexity. Each server requires separate credentials, configuration, and approval handling, leading to tool sprawl, fragmented governance, and limited visibility into costs. An MCP gateway resolves this by acting as a unified access layer in front of all upstream tool servers. Bifrost, an open-source AI gateway developed by Maxim AI, is designed specifically for this architecture.
Role of an MCP Gateway in Claude Code Environments
An MCP gateway functions as both an aggregation and governance layer between Claude Code and upstream MCP servers. It establishes a single connection to each tool server while exposing a consolidated /mcp endpoint to Claude Code. All tool invocations pass through this layer, where policies related to access control, observability, and routing are enforced before reaching the underlying systems.
In the absence of a gateway, each MCP server must be configured independently within Claude Code. The gateway model simplifies this by reducing multiple connections into a single interface, centralizing operational concerns such as authentication, auditing, budgeting, and tool filtering. The Model Context Protocol itself is an open standard that enables AI systems to dynamically discover and execute external tools. It was initially introduced by Anthropic in November 2024 and is now broadly adopted across AI platforms.
Limitations of Scaling MCP Servers Without a Gateway
As the number of connected MCP servers increases, several systemic issues arise:
- Configuration overhead: Each server requires its own configuration entry, transport setup, and credentials. Replicating this setup across teams introduces friction and inconsistency.
- Lack of centralized governance: Claude Code can access any tool exposed by connected servers without a unified policy layer to restrict usage by user, team, or project.
- Inefficient token usage: Every MCP server contributes its full set of tool definitions to the model context on each request. For example, five servers with thirty tools each result in 150 tool definitions being injected into every prompt. Anthropic has reported scenarios where this leads to 150,000 tokens per agent interaction.
An MCP gateway mitigates these challenges by introducing a centralized control plane.
Architecture of Bifrost’s MCP Gateway with Claude Code
Bifrost operates as both an MCP client and server. It connects upstream to MCP-compatible services such as filesystems, databases, GitHub, web search, internal APIs, Notion, and Slack, then aggregates these tools into a single /mcp endpoint. From Claude Code’s perspective, Bifrost appears as a single MCP server, while internally it manages multiple upstream connections.
In addition to tool aggregation, Bifrost also acts as a unified inference gateway, allowing Claude Code to route requests to non-Anthropic models without any client-side changes. This enables teams to use providers such as OpenAI, Azure OpenAI, Vertex AI, or open-weight models behind the same interface, while preserving Claude Code’s native workflow.
The MCP gateway in Bifrost supports three transport mechanisms:
- STDIO: Executes a subprocess and communicates via standard input and output, suitable for local tools.
- HTTP: Uses JSON-RPC to communicate with remote MCP servers, typically for cloud-hosted services.
- SSE: Maintains persistent connections through Server-Sent Events for streaming use cases.
Upon registering a new upstream server, Bifrost automatically discovers available tools and synchronizes them. Claude Code does not require updates when new tools are added. Additional configuration guidance is available in the Claude Code integration resource.
Setting Up an MCP Gateway with Claude Code
The setup process is minimal and assumes Node.js 18+ and an authenticated Claude Code environment.
Step 1: Launch Bifrost
Bifrost can be started locally using NPX or Docker:
npx -y @maximhq/bifrost
# or
docker run -p 8080:8080 maximhq/bifrost
Access the dashboard at http://localhost:8080. Deployment is also supported on Kubernetes, Docker Swarm, and bare metal environments.
Step 2: Register Upstream MCP Servers
Within the Bifrost dashboard, navigate to the MCP section and add each upstream server. Specify the connection type and provide the required endpoint or command. For HTTP-based servers, authentication headers such as API keys can be configured directly. Bifrost handles tool discovery and synchronization automatically. Detailed instructions are available in the MCP connection documentation.
Step 3: Define Virtual Keys with Scoped Permissions
Virtual keys serve as the primary governance mechanism. Each key defines which tools are accessible, along with constraints such as budgets, rate limits, and routing policies. Tool access is scoped at a granular level, enabling selective permissions within the same server. For example, a key may allow crm_lookup_customer while restricting crm_delete_customer. Refer to the virtual keys documentation for implementation details.
Step 4: Connect Claude Code to the Gateway
Add Bifrost as an MCP server in Claude Code:
claude mcp add --transport http bifrost http://localhost:8080/mcp
Verify the connection using /mcp within Claude Code. All permitted tools associated with the virtual key will be available. Future additions to Bifrost are automatically reflected without further configuration.
Governance and Access Control in Production
In production environments, unrestricted tool access is rarely acceptable. Bifrost enforces governance through two mechanisms:
- Virtual key scoping: Each key restricts access to a defined set of tools.
- MCP Tool Groups: Logical groupings of tools that can be assigned to users, teams, or services, enabling scalable permission management.
Every tool invocation is logged with metadata including tool name, server origin, input parameters, output, latency, associated virtual key, and the originating LLM request. This level of observability supports compliance requirements such as SOC 2 Type II, GDPR, HIPAA, and ISO 27001, as outlined in the Bifrost governance layer.
This centralized control plane is particularly critical in regulated industries, where access to sensitive systems must be tightly scoped, audited, and attributable to specific users or services. By enforcing policy at the gateway layer, teams can ensure consistent compliance across all MCP-connected tools without relying on per-server controls
For external deployments, Bifrost supports OAuth 2.1 with automatic client discovery and per-user identity mapping via OAuth authentication. This aligns with the MCP specification update released in March 2025.
Reducing Token Usage with Code Mode
A significant cost factor in multi-server MCP environments is context inflation caused by large tool catalogs. By default, all tool definitions are included in every request, increasing token usage substantially.
Bifrost’s Code Mode addresses this by representing MCP tools as a Python API. Instead of preloading all tool definitions, Claude Code dynamically invokes only the required tools for a given task. This approach minimizes context size, filters outputs before they reach the model, and consolidates multi-step workflows into a single execution cycle.
In environments with multiple MCP servers, this results in approximately 50 percent reduction in token usage and 30 to 40 percent improvement in latency. Additional insights on reducing token usage are detailed in the article on reducing token costs using Bifrost.
Recommended Operational Practices
Effective MCP gateway deployments typically follow these practices:
- Enable
enforce_auth_on_inferenceto ensure all requests are authenticated via virtual keys. - Deploy Bifrost behind HTTPS using a reverse proxy such as nginx or Cloudflare.
- Activate Code Mode when managing large tool catalogs to optimize cost and performance.
- Route both LLM and tool traffic through the same gateway to maintain unified observability and governance.
- Configure routing rules to support fallback across providers such as Vertex AI, AWS Bedrock, or Azure.
Bifrost introduces minimal latency overhead, measured at approximately 11 microseconds per request under sustained load of 5,000 RPS, ensuring it does not become a bottleneck.
Getting Started with Bifrost and Claude Code
Adopting an MCP gateway with Claude Code requires minimal configuration yet delivers significant operational benefits. Bifrost consolidates multiple tool connections into a single endpoint, introduces structured governance through virtual keys, and reduces token costs via Code Mode.
maximhq
/
bifrost
Fastest enterprise AI gateway (50x faster than LiteLLM) with adaptive load balancer, cluster mode, guardrails, 1000+ models support & <100 µs overhead at 5k RPS.
Bifrost AI Gateway
The fastest way to build AI applications that never go down
Bifrost is a high-performance AI gateway that unifies access to 15+ providers (OpenAI, Anthropic, AWS Bedrock, Google Vertex, and more) through a single OpenAI-compatible API. Deploy in seconds with zero configuration and get automatic failover, load balancing, semantic caching, and enterprise-grade features.
Quick Start
Go from zero to production-ready AI gateway in under a minute.
Step 1: Start Bifrost Gateway
# Install and run locally npx -y @maximhq/bifrost # Or use Docker docker run -p 8080:8080 maximhq/bifrost
Step 2: Configure via Web UI
# Open the built-in web interface
open http://localhost:8080
Step 3: Make your first API call
curl -X POST http://localhost:8080/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{ "model": "openai/gpt-4o-mini", "messages": [{"role": "user", "content": "Hello, Bifrost!"}] }'
That's it! Your AI gateway is running with a web interface for visual configuration…
To evaluate this architecture within your own infrastructure, book a demo with the Bifrost team.