Stacklok
Open-source platform ToolHive plus a paid enterprise edition for running MCP servers in isolation and securing AI agents' access, self-hosted.
What it does
Stacklok (Stacklok, Inc., a Delaware corporation) builds infrastructure that lets companies connect AI agents to tools and models in a controlled way. At its core is the open-source project ToolHive: it runs each MCP server in its own container (locally via Docker or Podman, or in Kubernetes), manages secrets and comes with a desktop app, a CLI and a Kubernetes operator. A registry server holds an approved catalogue of MCP servers, and the gateway combines several servers into one virtual MCP server (vMCP), with access policies, integration with your identity provider and audit logs. An LLM gateway is only part of the enterprise edition: it serves the OpenAI, Anthropic and Gemini APIs from one address, sets budgets per person or group and screens inputs for personal data and prompt injection.
Who it suits
Platform and security teams that want to allow many agents and MCP servers across the company without losing track of access and costs. Individual developers use the desktop app to run MCP servers locally in isolation.
Data protection for businesses
You run Stacklok in your own environment, locally or in Kubernetes; according to the vendor, customer data does not end up on its systems, and you decide on location and storage. According to its terms, however, the enterprise platform sends usage data to Stacklok, which the vendor uses to develop its products. Whether inputs are used for training depends on the model provider you connect. The vendor does not mention a data processing agreement. Stacklok Enterprise offers ready-made Okta and Entra ID integration; the open-source edition supports OIDC and OAuth.
Limits
Running it company-wide requires Kubernetes know-how. Pricing for the enterprise edition is only available on request. The vendor does not publish a privacy policy, a postal address or certifications such as SOC 2 on its website. It does not include a language model of its own.