Stateless MCP 2.0 reignites interest in safer, auditable AI agent tools
The new stateless Model Context Protocol specification makes AI agent tools easier to control and audit, reviving a standard that was eclipsed by more flexible but riskier approaches.
The Model Context Protocol (MCP), introduced by Anthropic in November 2024, has received its most significant update since launch with the release of the 2026-07-28 specification, often referred to as MCP 2.0 or Stateless MCP. The protocol defines a standard way to expose tools to LLM-powered agent frameworks. After a surge of interest through much of 2025, MCP was somewhat eclipsed by Skills, another Anthropic invention, which allowed agents to use a terminal and curl for more flexible interactions. However, the new stateless specification has reignited interest in MCP, as it addresses key risks associated with giving agents full shell and internet access.
Why it matters
This update signals a shift toward safer, more auditable AI agent architectures. As agents become more capable, the industry is grappling with the trade-off between flexibility and control. Stateless MCP offers a middle ground: tools that are easier to audit and control, and simple enough for smaller, locally-run models to drive effectively. This matters for developers building agentic systems, as it provides a standardized way to manage tool access without sacrificing safety.
This update signals a shift toward safer, more auditable AI agent architectures.
Agent sends request
LLM agent outputs a structured tool call with parameters
MCP server receives call
Stateless server processes the request without session context
Tool executes
Server runs the predefined function (e.g., API call, calculation)
Result returned
What you can learn from this
- Model Context Protocol (MCP) basics: MCP is a standard for exposing tools to AI agents. It works by defining a clear interface between the agent and external functions, like APIs or databases. As a learner, you should understand that MCP acts like a contract: the agent sends a request, the tool executes, and the result is returned. To practice, try building a simple MCP server that exposes a weather API or a calculator function, and connect it to a local LLM.
- Stateless vs. stateful design: The new stateless MCP means each tool call is independent, with no memory of previous calls. This simplifies auditing because you can log each request and response without tracking session state. For your projects, design your agent tools to be stateless where possible — it makes debugging and security reviews much easier.
- Auditability and control in agent systems: Giving an agent full shell access is risky because it can execute arbitrary commands. MCP tools are more auditable because they expose a limited, predefined set of actions. As a developer, you should always prefer tool-based interfaces over raw shell access for agents, and implement logging for every tool invocation.
- Smaller models and tool use: The excerpt notes that MCP tools are simple enough for smaller models running on a laptop. This is because tool calls are structured and predictable, reducing the cognitive load on the LLM. When building agent systems, test with smaller models first — if they can drive your tools, your design is likely robust and efficient.
- Trade-offs between flexibility and safety: Skills (terminal + curl) offer more flexibility but require strong models and introduce risk. MCP trades some flexibility for safety and auditability. As a learner, evaluate your use case: for high-stakes automation, prefer MCP-style tools; for exploratory tasks, you might accept more risk. Always document your threat model.
We teach this
Sources
- Stateless MCP has recaptured my interest (and inspired mcp-explorer and datasette-mcp) — Simon Willison
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report