Video summary

Did Anthropic finally fix MCP?

Main summary

Key takeaways

Technology

Tech focus: What changed in MCP (Model Context Protocol)

  • The speaker describes MCP as having been popular but overly complex—“bloated,” creating more problems than it solved.
  • A major update is now live: MCP moved to a “stateless core.”
  • The speaker frames this as effectively “MCP2” (or a renaming due to “destroyed” MCP branding), but also explicitly references the spec as:
    • MCP 2026-07-28 / “MCP 2.0”
  • The naming is criticized as confusing and dated.

Why stateless MCP matters (core technical benefits)

  • From stateful, persistent connections → request/response model

    • Previously, MCP required maintaining a bound connection between the client/agent and the MCP server, even when the agent might not call the tool.
    • This increased infrastructure and scaling complexity, including the need for sticky/session routing.

  • Stateless enables serverless/edge hosting

    • MCP servers can be deployed on serverless and edge infrastructure.
    • There’s no need for long-lived server state or session affinity.

  • Infrastructure becomes cheaper and simpler

    • The speaker argues MCP endpoints can be hosted behind a simple HTTP API (locally or in the cloud) using lightweight identifiers/keys rather than dedicated connections.

  • Better scalability + lower overhead

    • The speaker claims stateless MCP drastically reduces resource waste, making MCP viable as a real “integration” rather than an expensive/annoying experiment.

Concrete pain points from the old (stateful) MCP

  • The speaker describes a real-world overhead example:
    • Using Codeex with multiple MCP servers/sub-agents.
    • Each sub-agent can spawn multiple MCP servers, which previously required stateful connections and initialization behavior.
  • Result:
    • Large numbers of processes/connections
    • Noticeable local CPU/resource load
    • They reference macOS’s syspolicyd as an example of being triggered by many spawned services.
  • They also mention that initialization/handshake overhead could delay early interaction until MCP servers were ready—while Codeex mitigations exist, stateless MCP eliminates the underlying connection/state cost.

Security / authorization improvements mentioned

  • Alongside the stateless core, the update includes:
    • Hardened authorization
    • Graduated official extensions
    • Standardized extensions through apps and tasks
  • The speaker positions MCP as a way to constrain agent capabilities:
    • Instead of granting broad CLI access, MCP offers a self-describing set of endpoints/tools, limiting what the model can do.

Reviews / opinions / “is it fixed?” analysis

  • Overall stance: yes, it fixes the biggest architectural flaw (statefulness).
  • The speaker is excited to revisit MCP because it now resembles what they believe MCP should have been:
    • tools described via capabilities
    • simple calls to endpoints
  • Major caveat: compatibility/version fragmentation
    • Many MCP servers were built for the old stateful behavior and may not work with newer stateless clients.
    • Client-side MCP support varies across tools (e.g., Cloud Code, Codeex, Cursor, Pi).
    • Net effect: MCP may temporarily lose its “simple universal standard” status until the ecosystem converges.

“How-to / tooling” highlights mentioned

  • Official MCP site

    • Includes a visualizer to understand setup/behavior.

  • MCP Explorer / Python CLI by Simon Wilson (as described by the speaker)

    • Can query an MCP server and list available tools.
    • Example usage can be run via uvx without installing.
    • Demonstrates an MCP server for an agentic Mermaid demo, including tools such as:
      • “execute mermaid SDK code”
      • “render SVG”
      • “render PNG”
    • Mentions another MCP-related tool for data set integration, made feasible by stateless MCP.

  • Browserbase (sponsor example referenced)

    • Demonstrates using a Browserbase MCP with Claude/agents to interact with a website without APIs, using a real browser and dashboard monitoring.
    • This is more about browser-agent integration via MCP than about the MCP core change.

Named sources / main speakers

  • Main speaker (YouTuber)

    • Unnamed in subtitles.
    • Historically criticizes MCP; praises stateless MCP; discusses adoption risks.

  • Simon Wilson / Simon Willis

    • Cited as author of an article about “stateless MCP” and renewed interest.
    • Also credited with building MCP Explorer, the stateless Python CLI, and demo servers/tools.

Original video