← /editions
#148

The agent control plane goes open source

// Three open-source projects, covering auth for agents calling MCP servers, a governance layer, and agentic threat hunting, show the focus shifting from what agents can do to how they're controlled.

$ today_in_10_seconds

Three open-source projects hit Hacker News today, and all of them go after the same gap: who's allowed to do what when an agent acts on its own. One handles auth for MCP calls, one is a governance layer, and one uses agents to hunt threats. The plumbing around agents is finally getting as much attention as the agents themselves.

$ lead_story

Open-source auth for AI agents calling MCP servers

A builder has released an open-source auth layer for AI agents that call MCP servers and is looking for other builders to work on it. MCP made it easy for agents to reach tools. It did much less to settle how an agent proves who it is, whose authority it's acting under, and what it's allowed to touch. Many teams shipping MCP integrations are still solving that themselves with long-lived keys and shared tokens. The project is early, so read the code before you trust it. If you run agents against MCP servers in production, though, this is the problem to get ahead of.

// why_it_matters: An agent with tool access and no real identity is just an unaudited API key with initiative.

read more →

$ also_worth_attention

  • Bernstein: an open-source governance layer for AI agents

    Bernstein describes itself as an open-source governance layer for agents, a checkpoint between what an agent wants to do and what it's permitted to do. Auth answers “who is this agent.” Governance answers “should it do this, right now.” Expect those two layers to merge into one before long.

    link →
  • Argus: an open-source security platform that uses AI agents to hunt threats

    A Medium write-up highlights Argus, an open-source platform that uses AI agents for threat hunting, and calls it one of the most impressive open-source security projects of the year. Threat hunting suits agents well because the work is open-ended, multi-step, and tool-heavy: query, pivot, form a hypothesis, check again. Look at the repo before you buy the hype, and ask the question today's other two stories are circling: what is the hunter allowed to do once it finds something?

    link →

$ bigger_picture

The trend behind the headlines

For most of the agent boom, the conversation was about capability: can the agent reason, and can it use tools? Today's three launches are about control instead: identity, policy, and what autonomy looks like when you point it at your own infrastructure. Web apps went through the same shift when auth and permissions became standard infrastructure rather than something every team built from scratch. Here's my bet: the agent stack that wins won't be the one with the smartest model. It'll be the one a security team is willing to approve.

$ tool_of_the_day

The MCP specification's authorization section

Before you adopt any third-party MCP auth layer, including today's lead, read how the protocol itself handles authorization (it's built on OAuth 2.1). That's the baseline any tool in this space should be measured against.

explore The MCP specification's authorization section →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime