Stacklok
Stacklok provides an open, self-hosted, Kubernetes-native platform for enterprises to run and govern AI agents, enabling secure integration with sensitive systems and data. Its tools allow organizations to mix models and providers, maintain control over costs and data, and build agentic workflows using open source components. The mission is to ground AI in real-world systems while avoiding lock-in and supporting a vibrant community.
Facts
- HQ
- Seattle, Washington, US
- Sectors
- Software Development
- Founded
- 2023
- Employees
- 38
- Stage
- Series A
- Members see funding amounts. .
- Founders
- Craig McLuckie, Joe Beda
Is this your organization? Claim it
KubeCon 2026: Kubernetes Shifts Into Control Plane for AI Agents as Enterprises Prioritize Operating and Observing Agent Workloads
KubeCon 2026 showed a shift from Kubernetes as infrastructure for AI to Kubernetes as the control plane for AI agents, with industry focus moving from building agents to operating, scaling, and observing them as real enterprise workloads.
Scott Buchanan (Stacklok) outlines how teams should host MCP servers, recommending Streamable HTTP and Kubernetes for production
Scott Buchanan explains how to host MCP (Model Context Protocol) servers for teams, covering transport choices (Streamable HTTP vs stdio), deployment patterns (local, containerized, Kubernetes), and the governance needs—identity, segmentation, isolation, and observability—required for production use.
Skailar Hage: Enterprises should govern AI assistant skills with a centralized registry, not GitHub
Skailar Hage argues that enterprise AI assistant skills—code that changes agent behavior—are spreading unchecked like browser extensions and need centralized governance (a registry) for approval, versioning, and revocation rather than distribution via GitHub repositories.
Scott Buchanan: Enterprises Need Both MCP Gateways and LLM Gateways to Secure AI Agents
Scott Buchanan explains that safe enterprise AI needs two separate gateways: an MCP gateway that governs agents’ access to internal tools, data, and APIs, and an LLM gateway that governs what information is sent to models; both are required and policies must span the boundary between them.
Scott Buchanan: AI Agents Require Distinct Per‑Instance Identities Because Human and Service‑Account Models Fail at Scale
Scott Buchanan explains that AI agents need a distinct per-instance 'agent identity'—tied to their definition, linked to the human or service they act for, carrying scoped claims, inheritable by sub-agents and revocable—because existing human and service-account identity models each solve only part of the problem at scale.
Scott Buchanan: Enterprises Need Three Guardrails for AI Agents — Data, Model, and Infrastructure Controls
Scott Buchanan argues enterprises must build three guardrails for AI agents: a governed gateway to data, a gateway between agents and models, and infrastructure-level controls rather than relying on sandboxing alone.
Scott Buchanan (Stacklok) says generative AI eases code-level debt but won’t reduce structural production debt costs
Scott Buchanan argues that while generative AI can lower the cost of fixing feature- and code-level technical debt, it won’t make structural production debt (security, operations, recovery) cheaper to repay, and platform teams must still design resilient foundations rather than gamble on future AI fixes.
Skailar Hage announces Stacklok Enterprise adds per-gateway kill switches to immediately disconnect misbehaving MCP and AI gateways
Skailar Hage explains that organizations using MCP and AI gateways need an administrative kill switch so security teams can immediately stop misbehaving agents and verify traffic has been cut off, and announces Stacklok Enterprise now includes per-gateway kill switches in its admin settings.
Skailar Hage announces Stacklok Enterprise service accounts to give agents non‑human identities
Stacklok Enterprise introduces service accounts—non‑human identities for agents, pipelines, and scheduled automations—so autonomous workloads no longer run under departed employees' credentials, improving auditability, governance, and safe shutdowns.
Scott Buchanan: Agent Identity Must Be Set by the Runtime — Stacklok’s ToolHive and Mecatl Provide Verifiable Chains
Scott Buchanan argues that agent identity must be established at the runtime that launches the agent, not inferred later by a gateway, and describes Stacklok’s open-source runtimes (ToolHive and Mecatl) and enterprise platform as solutions that provide a verifiable chain of identity and governance.
Brian Dussault: Stacklok cuts AI costs — avoided OpenAI 272,000-token surcharge and found $3,900+ in monthly savings
Stacklok analyzed its AI gateway telemetry to find cost antipatterns—especially a pricing cliff in OpenAI’s models for requests over 272,000 input tokens—and by trimming unnecessary context and updating model selections they cut spending significantly and identified several thousand dollars in monthly savings potential.
Skailar Hage warns developer-focused AI governance breaks down as expired tokens and poor reconnection UX force knowledge workers to bypass controls
Skailar Hage describes how AI governance often assumes users are developers, so when connectors or tokens expire nontechnical knowledge workers encounter opaque errors and either stop using tools, share credentials, or route requests through untraceable accounts — behaviors that silently defeat governance controls. The piece calls attention to post-provisioning failures (reconnection and authentication UX) as a widespread, overlooked governance risk inside organizations.
Scott Buchanan explains the rise and meaning of 'agent harness' as discussed in LangChain, Microsoft, and OpenAI docs
Scott Buchanan explains why 'agent harness' has become a widespread term across major developer docs and forums, and walks through the most common questions about what it means and whether it matters.
Scott Buchanan: Own the Model, Harness, and Context—Especially the Tool Layer—in Your AI Agent Stack
Scott Buchanan argues that to build flexible, maintainable AI agents you should control the model, the execution harness, and the context—especially the tool layer—so you can swap models, tailor the loop to your domain, and retain agent memory and capabilities.
People
-
Craig McLuckie Executive Officer, Director, Founder and CEO
-
Joe Beda Founder, CTO