Automatic or Direct Tool Binding
Let the runtime automatically discover and bind allowed MCP tools into your LangGraph, LangChain, or Deep Agents workload at startup, or connect your own MCP client directly to the gateway URL.
Zelkor gives every agent a single MCP endpoint for Postgres, Qdrant, S3 object storage, the gVisor code sandbox, and external MCP servers—holding backend secrets at the gateway and never leaking your caller’s JWT to third parties.
When agents connect straight to databases, S3 buckets, and third-party APIs, every agent pod needs database passwords, storage keys, and broad outbound network access. Worse, nothing stops a prompt-injected model from running destructive SQL statements or sending internal user tokens to an external vendor.
Let the runtime automatically discover and bind allowed MCP tools into your LangGraph, LangChain, or Deep Agents workload at startup, or connect your own MCP client directly to the gateway URL.
Postgres: read-only queries and schema inspection with the caller’s tenant set on every transaction for Row-Level Security. Qdrant: document search and upsert with mandatory tenant filters injected automatically. Object Storage: tenant-prefixed file reads and writes on S3-compatible storage with a 256 KB payload cap and delete off by default. Code Sandbox: Python execution routed to the gVisor worker pool.
Register internal cluster services or external HTTPS MCP servers on the gateway. The gateway attaches the backend’s API key from a Kubernetes Secret and prevents external servers from spoofing built-in tool names.
When calling an external MCP server, the gateway strips the user’s internal Zelkor JWT and authenticates using only the secret configured for that backend.
workload:
agents:
mcpInject:
enabled: true
workspace:
tools:
enabled: true
postgresMCP:
enabled: true
qdrantMCP:
enabled: true