Stacklok builds open, self-hostable infrastructure for running AI agents and Model Context Protocol (MCP) servers inside the enterprise. It is aimed at platform and security teams who want employees to use agents and MCP tools without losing track of which tools each agent can reach, who is accountable for each action and where data goes.
The software runs on the customer’s own Kubernetes clusters, so data stays behind the customer’s firewall.
At A Glance
| Vendor | Stacklok |
|---|---|
| Category | MCP and AI agent infrastructure, DevSecOps |
| Identities covered | Human users mapped through identity providers, AI agents and their tool calls |
| Deployment | Self-hosted on customer Kubernetes infrastructure |
| Agentic AI | MCP Gateway, MCP Registry, ToolHive, LLM Gateway, Mecatl agent harness |
| Customers | TRM Labs, kt cloud, project44 |
| Founded | 2023, Seattle |
| Market classification | DevSecOps |
What Stacklok Does
Once agents can call tools, every MCP server becomes a new access path, and most organisations have no single view of who uses which one. Stacklok answers with an open architecture in which organisations can switch models and providers while keeping control of cost, data and direction. Every piece is built around Kubernetes and around identity: each agent action, tool call and model call is tied back to a responsible person or team, and one set of policies applies across AI workloads. An open source core means teams can start on a laptop and carry the same setup into shared clusters.
Products
ToolHive
ToolHive lets developers and platform teams run, secure and govern MCP servers and agent skills in one consistent way, from a laptop to a shared cluster. Individual developers use the desktop app, local and development work runs through the CLI, and multi-user environments use the Kubernetes Operator. Secrets management and container isolation are built in. Hooks for Claude Code and Cursor restrict MCP tool calls to servers that ToolHive manages, and the companion MCP Optimizer filters the tools those servers expose before they reach the client.
MCP Gateway
When every team wires agents to its own MCP servers, authentication and policy end up scattered. The MCP Gateway, a virtual MCP server (vMCP), gives each user a single endpoint for all their servers and handles authentication and authorisation centrally, using Amazon’s Cedar policy language. Endpoints can be segmented by identity group, and tool filtering trims what agents see, with a reported 85% or greater cut in input token costs.
MCP Registry
The ToolHive Registry Server builds a curated catalogue of approved MCP servers and skills and implements the official MCP Registry API, so employees pick from vetted tools instead of whatever they find online. Stacklok is a maintainer of the official MCP registry.
LLM Gateway And AI Gateway
Finance and platform owners use these gateways to set budgets and track AI spend by team, individual or project, while mixing frontier, hosted and on-premises models behind one control point.
Mecatl Agent Harness
Mecatl is an open source, cloud-native harness for running agents, letting customers rent models without being tied to a single model vendor.
How It Works
Everything is self-hosted on existing Kubernetes infrastructure, and agents are managed like any other Kubernetes workload. By design, “Stacklok does not host, process, transmit, or store customer data.”
| Component | What it does |
|---|---|
| ToolHive Runtime | Hosts the MCP server containers on the cluster. |
| Registry Server | Holds the catalogue of approved servers and skills. |
| Gateway | A secure proxy for inbound MCP traffic that protects context and credentials and applies organisational policy. |
| Interfaces | The UI and CLI discover, configure and run servers and set up connected AI clients automatically. |
| Observability | Audit logs and metrics, with a full trail of tool calls, are exported through standard telemetry tooling. |
Identities And Systems Covered
| Human users | Mapped through enterprise identity provider integrations. |
|---|---|
| AI agents | Each action, tool call and model call linked to a person or team. |
| Credentials | MCP servers no longer need their own OAuth, and users no longer keep API keys or service account tokens locally. |
| Systems | Internal services, data stores and SaaS tools reached through MCP servers. |
Integrations
| Area | Integrations |
|---|---|
| Identity providers | Okta, Entra ID |
| AI clients | Claude Code, Cursor |
| Observability | OpenTelemetry, Prometheus |
Company Background
| Founded | 2023. |
|---|---|
| Headquarters | Seattle, United States. |
| Leadership | Craig McLuckie (Co-founder and CEO), Joe Beda (CTO), Shanis Windland (COO) and Scott Buchanan (CMO). |
| Investors | Accel, Madrona and Bain Capital. |
Note: This overview is compiled from Stacklok’s publicly available information as of 3 November 2026. It is provided for information only. NHI Mgmt Group does not endorse any vendor, and nothing on this page represents the opinions of NHI Mgmt Group.