Why MCP Bridge

MCP Bridge Introduction 👋

What is MCP Bridge?

MCP Bridge is a self-hosted platform that turns any existing API into an AI-ready tool. Point it at a REST, GraphQL, SOAP, or gRPC endpoint, and it auto-generates Model Context Protocol (MCP) tool definitions that any LLM client can discover and invoke — no custom integration code required.

The Problem

Connecting LLMs to enterprise APIs today requires writing bespoke tool definitions for every endpoint, maintaining parameter mappings by hand, and rebuilding when APIs change. For organizations with dozens of internal services and hundreds of endpoints, this is a significant engineering tax that slows AI adoption.

The Solution

MCP Bridge eliminates that glue code, It:

1. Imports API schemas — Provide a schema via URL, paste content directly, or upload files. Supports OpenAPI (JSON/YAML), GraphQL introspection, WSDL, and gRPC (server reflection or .proto files). The system parses every operation automatically.

2. Generates MCP tools — Each API operation becomes a fully described MCP tool with typed input schemas, parameter mappings, and documentation — ready for any MCP-compatible client. Smart response processing defaults are auto-configured based on API type (e.g., array limiting for REST list endpoints, Relay connection unwrapping for GraphQL, flattening for SOAP).

3. Executes at runtime — When an LLM calls a tool, MCP Bridge validates inputs, maps parameters, handles authentication (Bearer, Basic, API Key, OAuth2, AWS Cognito SRP), and forwards the request to the backend API. Responses can be post-processed (filtered, transformed, aggregated) before flowing back through the MCP protocol — reducing token waste from large payloads.

4. Scales with Code Mode — For large APIs (100+ endpoints), Code Mode replaces the full tool catalog with 3 meta-tools, cutting context window usage by ~98%. The LLM discovers tools on demand and orchestrates calls via JavaScript in a secure sandbox.

Key Capabilities

Capability	Details
API Types	REST (OpenAPI 3.x / Swagger 2.0), GraphQL (introspection), SOAP (WSDL), gRPC (server reflection / .proto files)
Schema Input	Three methods: fetch from URL, paste content (JSON/YAML/XML/proto), or upload files (.json, .yaml, .wsdl, .proto — multiple .proto files supported for gRPC)
Authentication	None, Bearer token, Basic auth, API key (header/query), OAuth2 (Client Credentials, ROPC), AWS Cognito SRP (USER_SRP_AUTH flow)
Web UI Auth	Two login modes: (1) self-hosted login page with Cognito SRP — no external redirect, credentials validated server-side; (2) external OIDC provider redirect (OAuth2 Authorization Code + PKCE) — works with any provider (Cognito, Entra ID, Keycloak, Auth0, Okta). Both use encrypted session cookies
MCP Transport	Streamable HTTP (/mcp endpoint), with optional OAuth2 resource server protection
Tool Management	Enable/disable individual tools, edit names and descriptions, customize parameter mappings, configure response post-processing
Response Processing	Per-tool response post-processing to filter, transform, or aggregate API responses before they reach the MCP client. Two modes: declarative rules (unwrap, select, exclude, limit, sort, flatten, aggregate) or custom JavaScript. Smart defaults are auto-generated during schema import — REST GET array endpoints get limit(100), REST wrapper objects get unwrap + limit(100), GraphQL Connection queries get unwrap("data.<name>.edges") + limit(100), SOAP operations get flatten(1), gRPC server-streaming methods get limit(100). Defaults are fully customizable in the Tool Editor
Code Mode	3 meta-tools (search_tools, get_tool_schema, execute_code) replace N tool definitions; Boa JavaScript sandbox with 30s timeout
Reliability	Per-API rate limiting (token bucket), exponential backoff with jitter, configurable retry policies
Observability	Execution logs with input/output token estimates, API health checks, per-tool token cost badges, analytics dashboard
Analytics	Tabbed dashboard (Overview, Tools, APIs, Token Usage, Errors) with configurable time ranges (1h / 24h / 7d / 30d). Overview: latency distribution, latency over time, throughput, peak usage heatmap. Tools: per-tool metrics, slowest executions, unused tools. APIs: per-API metrics, availability & failure streaks. Token Usage: summary, per-tool breakdown, cost trend, most expensive executions. Errors: rate trends, category breakdown, per-tool error rates, error correlation
Search	Hybrid search combining full-text search, trigram fuzzy matching, and (optional) vector-based semantic similarity. Semantic search uses pgvector with HNSW indexing and 384-dimensional embeddings for meaning-aware tool discovery

Who It's For

• Platform teams that want to expose internal APIs to AI agents without writing MCP adapters for each service.

• AI engineers building agents that need to call enterprise APIs and want a managed tool layer with auth, rate limiting, and observability built in.

• Organizations adopting MCP as a standard and looking to bridge their existing API portfolio to LLM clients quickly.