<- all tokdocs

Is MCP Dead? Code Execution With MCP and the 77% Token Claim

By , founder of Mavric and Apso

View on TikTok ->

A TikTok from @ducktate.com says "traditional MCP is pretty much dead" and that a new method called code execution with MCP saves up to 77% in usage. The 77% figure is real, but it comes from a single benchmark. Both companies that published the technique describe it as a way to use MCP, and the MCP spec shipped a general availability release three months before the video.

Is MCP dead?

No. The two primary sources behind the technique say the opposite.

Cloudflare's post by Kenton Varda and Sunil Pai, published September 26, 2025, is titled "Code Mode: the better way to use MCP." It states that "MCP is still useful" because it "provides a uniform way to connect to and learn about an API."

Anthropic's engineering post, published November 4, 2025, is titled "Code execution with MCP: Building more efficient agents." It describes the change as a presentation layer: "a solution is to present MCP servers as code APIs rather than direct tool calls."

In both, the agent still connects to MCP servers. The difference is that the model writes code against generated API bindings instead of making one tool call per step. Even the AIMultiple article shown in the video labels its two test arms "Regular MCP" and "Code execution MCP."

The protocol is also still shipping. MCP published its 2026-07-28 specification on July 28, 2026 as a general availability release. The announcement reports close to half a billion downloads a month across the Tier 1 SDKs, with the TypeScript and Python SDKs each past one billion cumulative downloads.

What is code execution with MCP?

In regular MCP tool calling, the definitions for every tool on every connected server load into the model's context window up front, and the model makes tool calls one at a time. AIMultiple's article names that up-front loading as the defining property of "Regular MCP."

With code execution, the model writes a short program that calls the MCP tools, and a sandbox runs it. Cloudflare's reasoning is that "LLMs are better at writing code to call MCP, than at calling MCP directly," because models have seen far more real code than synthetic tool-calling data.

The video explains the idea with a grocery store analogy about repeated phone calls. That analogy covers round-trip overhead but not the tool-definition loading that drives most of the saving.

Does code execution with MCP really cut token use by 77%?

On one benchmark, yes. AIMultiple's article "Code Execution with MCP: A New Approach to AI Agent Efficiency," by Şevval Alper and updated September 9, 2026, reports these averages:

  • Total tokens: 775,197 for regular MCP against 175,081 for code execution, a 77.4% reduction.
  • Average input tokens: 15,417 against 3,310, a 78.5% reduction.
  • Average latency: 9.66 s against 10.37 s, about 7% slower.
  • Success rate: 100% for both.

The video calls this "a bunch of different benchmarks." It is one benchmark: one model (GPT-4.1), one MCP server (Bright Data's MCP Server in Pro Mode), two task types and 50 runs per approach. AIMultiple lists its own limits: only two task types, only one model, and success rates that may fall on tasks where code quality matters more.

What does it actually save in dollars?

Less than 77%. The same table, visible in the video, shows output tokens rising 120%, from 87 to 192 on average and from 4,345 to 9,585 in total. The agent writes more because it is generating code.

Output tokens cost more than input tokens. OpenAI prices gpt-4.1 at $2 per million input tokens and $8 per million output tokens. Applied to the table's totals:

  • Regular MCP: $1.5417 input plus $0.0348 output, for $1.5765.
  • Code execution: $0.3310 input plus $0.0767 output, for $0.4077.

That is a 74.1% cost reduction. The speaker said "token usage," so the 77% claim is accurate as stated. For a budget, use 74.1% for this model and task mix, and expect the gap to grow with how much code the agent writes.

Who invented code execution with MCP?

Cloudflare and Anthropic, not AIMultiple. The video presents AIMultiple as the source of the method, but Cloudflare published Code Mode on September 26, 2025 and Anthropic published its post on November 4, 2025. AIMultiple's article, dated September 9, 2026, measured the technique about ten to eleven months later.

Anthropic's own example is larger than the number in the video. It describes a workflow that reads a meeting transcript from Google Drive and attaches it to a Salesforce lead, where "This reduces the token usage from 150,000 tokens to 2,000 tokens, a time and cost saving of 98.7%." That is one illustrative workflow, not a multi-run benchmark.

Which Cloudflare sandbox do you need?

The video says you need "some sort of sandbox environment" and shows the Cloudflare Sandboxes product page, headlined "Secure code execution for AI agents and developer tools." That product runs on Cloudflare Containers. Its cards read "Fully isolated," "Fast spin-up" and "Bring your own images," and it includes a built-in code interpreter through runCode().

Cloudflare's Code Mode post uses a different runtime. It runs generated code in V8 isolates on Workers and argues that "isolates are far more lightweight than containers," with a fresh isolate per snippet. Both are Cloudflare code execution options, with different isolation models and costs, and the video treats them as one product.

The speaker's warning to set the sandbox up carefully matches Anthropic's own caveat: "Running agent-generated code requires a secure execution environment with appropriate sandboxing, resource limits, and monitoring."

The speaker also says, "I haven't tested this out personally." The clip contains no terminal, code or test run, only scrolls through the AIMultiple article and the Cloudflare page.

Bottom line

MCP is not obsolete. Code execution is a way of calling MCP servers that Cloudflare and Anthropic published in 2025. The 77.4% token cut is real for one GPT-4.1 benchmark, and the cost saving on that benchmark is 74.1% because output tokens rise 120%.

Sources

Published October 26, 2026. Based on a TikTok by @ducktate.com. Full video notes (markdown)