gateway.karada.ai/mcp).
Dual Audience & Platform Architecture:
- For Developers & Enterprises (B2C/B2B Product): Set up configurations once on Karada and use seamlessly across multiple harnesses (Cursor, Claude Code, Windsurf) while enforcing organizational security standards and tool access policies across the entire team.
- For MCP Creators & API Teams (Distribution Play): Publish to our registry to get your tools directly in front of users using our universal gateway across Cursor, Claude Code, Windsurf, and enterprise agent fleets.
How the Gateway Works
When your AI client sends atools/list request, the gateway concurrently queries all connected upstreams, merges the results with namespace prefixes (e.g., stripe__createCustomer, github__listRepos), and returns a single unified tool catalog. Tool calls are automatically validated and routed to the correct upstream service.
Multi-Gateway Architecture
You can create multiple gateways to organize tools by team or security context—for instance, one gateway for backend engineering tools and another for finance and analytics APIs. Each gateway provides:- Dedicated Authentication Token: Scoped per gateway endpoint
- Independent Server Inventory: Attach only the MCP servers needed for that context
- Granular Tool Toggles: Enable or disable specific tools without disconnecting the upstream server
- Organizational Standards: Enforce tool versions and access policies across all team members
Connecting MCP Servers
From the Tools tab in your Gateway dashboard, connect MCP servers in two ways:- From the Registry: Browse verified servers in Karada’s curated MCP registry and connect with one click.
- Custom URL: Provide any Streamable HTTP MCP server endpoint. Karada introspects the upstream to discover available tools and parameter schemas automatically.
Automated Tool Namespacing
To prevent naming collisions when multiple servers define generic tools likesearch or create_user, Karada prefixes tool identifiers with the server namespace (e.g., github__search, notion__search).
Server-Side Credential Delegation
Many upstream MCP servers require API keys or OAuth tokens. The Gateway handles server-side credential delegation:- Define header names and secret values in the dashboard
- Karada injects credentials directly into upstream HTTP requests at the gateway edge
- Secrets are encrypted at rest using AES-256-GCM
- Secrets are shielded from the AI client’s context window
Autonomous AI Agent (karada_agent)
The Gateway includes a built-in autonomous AI agent (karada_agent) capable of executing multi-step workflows across your connected tools without bloating your AI assistant’s prompt context:
- Analyzes the merged tool catalog across all connected MCP servers
- Plans a step-by-step sequence of operations
- Executes calls server-side, passing output context between tools
- Returns a complete execution trace and final synthesized answer, saving up to 95% of client context tokens.
Gateway Configuration Modes
Standard vs Enhanced Mode
- Standard: Core multiplexing, tool routing, namespace isolation, and credential delegation.
- Enhanced: Adds dynamic real-time tool discovery, autonomous task execution (
karada_agent), and higher throughput limits.
Dynamic Discovery
When enabled, the gateway re-introspects connected upstreams on eachtools/list request to surface newly added tools immediately. For maximum speed on stable schemas, disable dynamic discovery to serve cached tool manifests instantly.