DOCSEARCHbyAlgolia
Submit Docs

Stacklok Enterprise

website·Indexed 35 days ago

https://docs.stacklok.com

A governance platform for AI infrastructure providing centralized control over LLMs and MCP servers. Key components include the AI Gateway for model routing, budget management, and PII/prompt injection screening, as well as the Connector Gateway for managing MCP servers. The platform features enterprise authorization, OIDC and SCIM identity integration, and a dedicated administrative console and CLI.

Tokens
347.6K
Snippets
1.5K
Records
2.4K
Agent score
67%

What's inside Stacklok

  1. Overview of the Filesystem MCP server

    The Filesystem MCP server allows AI agents to read and write files on the local filesystem. It is particularly useful for accessing files outside the current working directory, operating in headless environments, or demonstrating MCP capabilities. The server uses stdio transport and does not require network connectivity.

  2. Overview of Virtual MCP Server (vMCP)

    Virtual MCP Server (vMCP) is a ToolHive feature that acts as an aggregation proxy, consolidating multiple backend MCP servers into a single unified interface. This allows clients to connect to one endpoint to access all backend tools rather than configuring connections to each server individually. vMCP provides multi-server aggregation, automatic tool conflict resolution via namespacing, centralized authentication, and support for composite workflows with parallel execution and approval gates.

  3. Overview of the Fetch MCP server

    The Fetch MCP server (implemented as GoFetch by Stacklok) is a lightweight tool that retrieves data from websites to provide real-time context to AI agent workflows. It is a high-performance, secure Go implementation of the original Model Context Protocol (MCP) fetch server.

  4. Overview of ToolHive for MCP Server Management

    ToolHive is an enterprise-grade open-source (Apache 2.0) platform designed for running, discovering, securing, and operating Model Context Protocol (MCP) servers and agent skills. It allows users to run any MCP server regardless of the technology stack by containerizing them, providing a consistent management layer across local and cloud environments.

  5. Overview of MCP Optimizer for token reduction

    The MCP Optimizer is an intelligent intermediary between AI clients and MCP servers designed to reduce token usage and improve tool selection accuracy. Instead of exposing every available backend tool to the model (which consumes significant context window/tokens), the optimizer exposes only two lightweight primitives: find_tool for semantic search and call_tool for routing.

    Workflow:

    1. The assistant calls find_tool with keywords extracted from the prompt.
    2. The optimizer returns the most relevant tools (default is up to 8).
    3. Only these specific tools and their descriptions are included in the model's context.
    4. The assistant uses call_tool to execute the task.

  6. Overview of the Kubernetes MCP server

    The Kubernetes MCP server (based on the MKP project) allows LLM-powered applications to interact with Kubernetes clusters to manage and automate tasks. It provides tools to list, get, and apply Kubernetes resources via the Kubernetes API.
    {
  "name": "io.github.stacklok/k8s",
  "image": "ghcr.io/stackloklabs/mkp/server:0.4.3",
  "tools": ["list_resources", "get_resource", "apply_resource"]
}

  7. Overview of MCPServerEntry for remote MCP servers

    MCPServerEntry is a zero-infrastructure catalog entry used to declare a remote MCP server endpoint for Virtual MCP Server (vMCP) discovery and routing. Unlike MCPRemoteProxy, it does not create pods, services, or deployments. It is used when a remote server is trusted, handles its own authentication, or when the overhead of a proxy is unnecessary.

    MCPServerEntry resources must be part of an MCPGroup. When vMCP starts in discovered mode, it queries the referenced group and connects directly to the remote URL. It includes built-in SSRF protection via URL validation to block internal and metadata endpoints.

  8. Overview of ToolHive for MCP server management

    ToolHive is the open-source core for running, securing, and governing Model Context Protocol (MCP) servers and agent skills. It supports deployment across desktop applications, CLI, and Kubernetes.

  9. Overview of the ToolHive embedded authorization server in Kubernetes

    The embedded authorization server allows ToolHive to handle the full OAuth flow, including redirecting users to an upstream identity provider (IdP). This is specifically useful for MCP servers or remote MCP endpoints that require Authorization: Bearer tokens but do not have a direct federation relationship with the identity provider (e.g., GitHub, Google Workspace, or Atlassian).

  10. Overview of Cedar authorization in Kubernetes

    Cedar authorization works alongside authentication to determine what an authenticated client is allowed to do. When using the embedded authorization server, Cedar policies can evaluate upstream identity provider claims, such as group membership. Authorization is applied to MCPServer, MCPRemoteProxy, or VirtualMCPServer resources.

  11. Overview of the vMCP Embedded Authorization Server

    The embedded authorization server runs an OAuth authorization server within the vMCP process. It redirects users to upstream identity providers (e.g., GitHub, Google, Okta) for authentication, stores the resulting upstream tokens, and issues its own JWTs for the vMCP to validate. This allows backend MCP servers to call external APIs on behalf of users even when no federation relationship exists between the identity provider and those services. It also supports OAuth 2.0 Client ID Metadata Documents (CIMD) and Dynamic Client Registration (RFC 7591).

  12. Overview of ToolHive UI for MCP Management

    The ToolHive UI is a desktop portal designed to simplify the setup and management of Model Context Protocol (MCP) servers on a local machine. It allows users to discover servers via a built-in registry, run them securely, and connect AI clients (such as Claude or Cursor) without manually managing Docker commands or configuration files. It can connect local AI clients to the broader ToolHive platform, including servers managed by the Kubernetes Operator and catalogs from a ToolHive Registry Server.