MCPJungle

MCPJungle acts as a centralized MCP-compliant proxy that consolidates multiple upstream MCP servers (stdio, SSE, HTTP transports) into a single gateway endpoint. Agents connect once to MCPJungle's /mcp endpoint instead of configuring individual server connections; the gateway internally maintains persistent connections to all registered servers, multiplexes tool discovery requests, and routes tool invocations to the correct upstream server based on canonical naming (server__tool format). This eliminates N×M configuration complexity where N agents must each configure M servers.

MCPJungle manages connections to upstream MCP servers across three transport types: stdio (local process spawning), SSE (Server-Sent Events over HTTP), and HTTP (bidirectional JSON-RPC). The gateway maintains a transport abstraction layer that handles protocol-specific connection lifecycle (spawn/connect/reconnect/disconnect), message serialization/deserialization, and error recovery. Each registered server's transport type is persisted in the database; MCPJungle automatically handles reconnection logic with exponential backoff for failed connections, enabling heterogeneous server ecosystems where some servers are local processes and others are remote HTTP endpoints.

MCPJungle supports two deployment modes: development mode (single-process, in-memory state, no authentication) and enterprise mode (distributed, persistent database, authentication/authorization, observability). Development mode is suitable for local testing and prototyping; enterprise mode adds production-grade features including database persistence, access control, audit logging, and metrics collection. Mode is selected at startup via configuration; switching modes requires database migration.

MCPJungle provides multiple deployment options: Docker containers (with docker-compose for local development and production), standalone binaries (Linux, macOS, Windows), and Kubernetes-ready configurations. Production deployments support environment variable configuration, database connection pooling, TLS/mTLS for upstream server connections, and horizontal scaling behind a load balancer. Docker images are published to registries; binaries are built via GoReleaser for multiple platforms.

MCPJungle provides a native Go client library that enables Go applications to programmatically manage MCPJungle (register servers, manage tools, define access policies) and invoke tools through the gateway. The client library wraps the HTTP API with type-safe Go methods, handles authentication, and provides structured error handling. This enables Go-based infrastructure automation, monitoring systems, and custom management tools to integrate with MCPJungle without writing HTTP requests manually.

MCPJungle aggregates tool definitions from all registered upstream servers via the MCP tools/list protocol, applies a canonical naming scheme (server__toolname) to prevent collisions, and exposes the merged catalog through a single tools/list endpoint. The gateway caches tool definitions in its database with server provenance metadata, enabling fast discovery without querying all upstream servers on every request. When agents invoke tools, MCPJungle parses the canonical name to route the invocation to the correct upstream server, transparently stripping the server prefix before forwarding to the target server.

MCPJungle intercepts tool invocation requests (tools/call) from agents, parses the canonical tool name (server__toolname) to identify the target upstream server, and routes the invocation to that server while preserving session context and request metadata. The gateway maintains per-session state including authentication tokens, request IDs, and invocation history, enabling stateful tool interactions where multiple tool calls within a session share context. Tool results are returned to the agent with metadata about execution time, server, and any errors, enabling observability and debugging.

MCPJungle supports organizing tools into logical groups (e.g., 'file-operations', 'web-search', 'database-admin') and filtering which tools are available to specific agents or users. Tool groups are defined in the database and can include tools from multiple servers; agents can request a filtered tool list (e.g., tools/list?group=file-operations) to see only tools in that group. This enables fine-grained access control and reduces cognitive load for agents that only need a subset of available tools.

MCPJungle's enterprise mode provides server-level access control via allowlists, enabling administrators to specify which agents or users can access which upstream servers. Access control is enforced at the gateway level before tool invocations are routed to upstream servers. The system supports multiple authentication methods (API keys, OAuth, mutual TLS) and integrates with the HTTP API and CLI for managing access policies. Access decisions are logged for audit purposes.

MCPJungle aggregates prompt templates from upstream MCP servers (via the MCP prompts/list protocol) and exposes them through a unified prompts/list endpoint, similar to tool aggregation. Prompts are cached in the database with server provenance metadata. Agents can request prompt templates by canonical name (server__promptname), and MCPJungle retrieves the template from the upstream server, applies any variable substitution, and returns the rendered prompt to the agent. This enables agents to access reusable prompt templates from multiple servers without reconfiguring.

MCPJungle collects detailed metrics about tool invocations, server connections, and gateway performance, including execution time, success/failure rates, error types, and per-server statistics. Metrics are exposed via Prometheus-compatible endpoints and can be scraped by monitoring systems like Prometheus or Datadog. The gateway also maintains structured logs of all tool invocations, server connections, and access control decisions, enabling debugging and audit trails. Observability data includes request IDs, timestamps, agent identifiers, and execution context.

MCPJungle provides a comprehensive CLI (mcpjungle command) for managing servers, tools, prompts, and access control without using the HTTP API. The CLI supports commands for server registration/deregistration, tool enable/disable, tool group management, access control policy definition, and server lifecycle operations (start, stop, status). CLI commands interact with the MCPJungle HTTP API and persist configuration to the database. This enables infrastructure-as-code workflows where server configurations can be version-controlled and deployed via CI/CD pipelines.

MCPJungle exposes a comprehensive HTTP API (REST + JSON) for all management operations: server registration/deregistration, tool enable/disable, tool group management, access control policy definition, and server status queries. The API supports authentication via API keys, OAuth, or mutual TLS. All API responses are JSON with consistent error handling and status codes. The API enables programmatic integration with external systems (CI/CD, monitoring, configuration management) and supports both synchronous and asynchronous operations.