<- all tokdocs
2026.10.26@ducktate.com

The video says MCP is dead; both primary sources for the technique say code execution runs on top of MCP, and the MCP spec reached GA on 2026-07-28

AI
Watch on TikTok

View on TikTok ->

A 79.30 s vertical clip, 720x1280 SDR at 30 fps, QuickTime/MP4 container (mov,mp4,m4a,3gp,3g2,mj2), H.265/HEVC Main profile video at 541,261 bps with 2,379 frames, HE-AACv2 audio at 48,097 bps, 44.1 kHz stereo, 79.297007 s audio against 79.300000 s video, total 5,908,960 bytes at an overall 596,111 bps, format id bytevc1_720p_596111-1; posted at epoch 1793047283, which is 2026-10-26T20:41:23Z, by handle ducktate.com (uploader id 7665287672349737997) under the channel name "Chris Ducktate", audio credited as "original sound" by Chris Ducktate; the description is 40 characters and duplicates the title exactly, "Apparently MCP is obsolete for AI AGENTS"; metadata was captured at 2026-10-26T23:03:50Z, roughly two hours and twenty minutes after posting, showing 1,207 views, 41 likes, 2 comments, 15 reposts and 33 saves. I read 19 frames at 2 s spacing (001, 004, 007, 012, 019, 023, 024, 025, 026, 027, 028, 029, 030, 031, 032, 033, 034, 036, 040) and the full 21-segment, 315-word transcript. Frames 001 through 023 and 027 through 030 show a man in a white snapback cap with a red square patch, clear-framed glasses and a light blue patterned polo, seated at a wood table in a room with a navy accent wall, louvered white closet doors and a ceiling fan, holding a black speckled enamel mug reading "REACT SUMMIT"; captions are burned in one word at a time in white bubble type, reading "MCP IS DEAD?" at frame 001 (0 s), a yellow arced "TOKEN HACKING" title card over "WELCOME" at frame 004 (6 s), then "BURN", "CALL", "MCP", "ACTUALLY", "UP", "USAGE", "TO", "TOOLS", "TOMORROW". Frames 024, 025 and 026 (46 s, 48 s, 50 s) cut to a phone screen recording of an AIMultiple article, logo rendered as one word "AIMultiple", with a sticky "In This Article" bar linking "Code execution with MCP vs regular MCP" and a dark table listing Success Rate 100% / 100% / Same, Avg Latency 9.66s / 10.37s / +7%, Avg Input Tokens 15,417 / 3,310 / -78.5%, Avg Output Tokens 87 / 192 / +120%, Total Input Tokens 770,852 / 165,496 / -78.5%, Total Output Tokens 4,345 / 9,585 / +120%, Total All Tokens 775,197 / 175,081 / -77.4%, followed by "Download as CSV" and body text beginning "We compared two approaches for building AI agents that interact with external tools via the MCP:" with bullets for "Regular MCP: Traditional approach where all tool definitions are loaded into the model's context window" and "Code execution MCP: Novel approach where". Frames 031 through 034 (60 s to 66 s) cut to the Cloudflare Sandboxes product page, showing a dashed "Sandboxes" chip, the headline "Secure code execution for AI agents and developer tools", the subhead "Run untrusted code in isolated environments. Give your AI agents, code interpreters, and developer tools a secure place to execute code, install packages, and interact with the filesystem.", a red "Start building for free" button and a "View docs" button, then cards for "Fully isolated", "Fast spin-up" and "Bring your own images", then a "How Sandboxes work" section with "SDK-driven control", "Built on Containers", "Real-time streaming" and "Code interpreter built-in". No browser address bar, no URL, no terminal, no editor and no running code appear in any frame I read, and no demo of the technique is shown anywhere in the clip. The spoken argument is that traditional MCP tool calling wastes tokens the way repeated phone calls from a grocery store waste time, that the replacement is "code execution with your MCP", that an AIMultiple benchmark shows up to 77% lower token usage at slightly higher latency, and that implementing it requires a sandbox such as Cloudflare's, with the speaker stating at 71.36 s that he has not tested it himself. The transcript ends cleanly: the last segment runs 78.40 s to 79.14 s, inside the 79.297007 s audio stream, so there is no Whisper hallucination past the end of audio and every quoted line below is genuine.

The central claim is wrong: neither primary source says MCP is obsolete, and the spec went GA three months before this video

The video opens at 0.00 s with "Apparently using a traditional MCP is pretty much dead, but this new method is going to save you like up to 77% in usage." Frame 001 burns the caption "MCP IS DEAD?" across the speaker's chest, and the TikTok description reads "Apparently MCP is obsolete for AI AGENTS."

The two primary publications behind the technique say the opposite. Cloudflare's post, written by Kenton Varda and Sunil Pai and published on September 26, 2025, is titled "Code Mode: the better way to use MCP" and states directly that "MCP is still useful", arguing its value is that it "provides a uniform way to connect to and learn about an API" regardless of the implementation behind it. Anthropic's engineering post, published November 4, 2025, is titled "Code execution with MCP: Building more efficient agents" and frames the change as a presentation layer, writing that "a solution is to present MCP servers as code APIs rather than direct tool calls." Both describe agents that still connect to MCP servers. The difference is that the model writes code against generated API bindings instead of emitting one tool call per step.

The AIMultiple article the video actually puts on screen agrees. Its own body text, visible in frame 026 at 50 s, defines the two arms of its comparison as "Regular MCP" and "Code execution MCP", both of them MCP, and its architecture description routes the code execution agent through an MCP Client Session.

The protocol itself is not winding down. The Model Context Protocol published its 2026-07-28 specification revision on July 28, 2026, announced as a general availability release: "Today, we're officially pushing the release button on the next version of the MCP specification, 2026-07-28, along with the SDKs that will allow you to start building clients and servers right away." That post 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. A specification that ships a GA revision with a stateless protocol core, header-based routing and a formal extensions framework three months before this video is not obsolete.

The 77% figure is real and traceable, but it comes from one benchmark, not "a bunch of different benchmarks"

At 45.68 s the video says "So here's an article here from AI Multiple and it has a bunch of different benchmarks." At 49.22 s it continues: "And you can see that the latency does go up, but if you actually look at the amount of token usage, it goes way down, up to 77%."

The number checks out against the source on screen. AIMultiple's article, "Code Execution with MCP: A New Approach to AI Agent Efficiency" by Şevval Alper, updated September 9, 2026, publishes the exact table the frames show, and its Total All Tokens row reads 775,197 for regular MCP against 175,081 for code execution, a difference of -77.4%. Dividing 175,081 by 775,197 gives 0.2259, confirming the stated 77.4% reduction. The latency claim also holds: 10.37 s against 9.66 s is a 7.3% increase, which the table rounds to +7%.

The description of the evidence does not hold. AIMultiple ran a single benchmark, not several. Its methodology section specifies one model, GPT-4.1, one MCP server, Bright Data's MCP Server in Pro Mode, two distinct task types, and 50 runs per approach, with the generated code executed in a sandboxed code executor using an async call_tool() function. The run count is verifiable from the table's own arithmetic: 770,852 total input tokens divided by 15,417 average input tokens equals 50.0, and 165,496 divided by 3,310 equals 50.0. AIMultiple lists three limitations itself: only two task types were tested, only one model was tested so results may vary elsewhere, and success rates may decline on tasks where code quality matters more. A single-model, two-task, 100-run study is a useful data point and is not "a bunch of different benchmarks."

The same table on screen shows output tokens rising 120%, and the real dollar saving is 74.1%, not 77.4%

Frames 024, 025 and 026 display two rows the narration skips entirely. Avg Output Tokens go from 87 under regular MCP to 192 under code execution, and Total Output Tokens go from 4,345 to 9,585, both marked +120%. Dividing 192 by 87 gives 2.207, confirming the +120% figure. The agent writes more tokens because it is now generating code rather than emitting short tool-call arguments.

That matters because output tokens cost more than input tokens. OpenAI's model page for gpt-4.1, the model AIMultiple benchmarked, prices input at $2 per million tokens and output at $8 per million. Applying those prices to the table's own totals gives regular MCP 770,852 input tokens at $1.5417 plus 4,345 output tokens at $0.0348, for $1.5765, against code execution's 165,496 input tokens at $0.3310 plus 9,585 output tokens at $0.0767, for $0.4077. The cost reduction is 74.1%, about three and a third points below the 77.4% token reduction the video quotes.

The 77% claim is accurate as stated, because the speaker says "token usage" at 49.22 s and not cost. Anyone taking the number to a budget forecast should use 74.1% for this benchmark's model and task mix. Output token growth is the term that erodes the saving, and it grows with how much code the agent writes.

The video credits AIMultiple for a method Cloudflare and Anthropic published roughly a year earlier

At 33.90 s the video says "This new method is actually called code execution with your MCP", then at 45.68 s presents AIMultiple as the place it comes from. The chronology runs the other way.

Cloudflare published "Code Mode: the better way to use MCP" on September 26, 2025, making the core argument that "LLMs are better at writing code to call MCP, than at calling MCP directly", on the reasoning that models have seen enormous volumes of real code and comparatively little synthetic tool-calling data. Anthropic published "Code execution with MCP: Building more efficient agents" on November 4, 2025, and announced it the same day on X. AIMultiple's benchmark article carries an updated date of September 9, 2026, roughly ten to eleven months after both. AIMultiple measured the technique. It did not introduce it.

The attribution gap also costs the video its strongest number. Anthropic's post reports a single workflow, reading a meeting transcript from Google Drive and attaching it to a Salesforce lead record, 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 rather than a benchmark across runs, and it should be read as such, but it is the figure that made the technique visible and it is 21 points larger than the one the video cites.

One naming note. The Whisper transcript renders the publisher as "AI Multiple" in two words at 45.68 s. The logo in frames 024, 025 and 026 reads "AIMultiple" as one word, with "AI" in blue and "Multiple" in grey, matching the site's own branding at aimultiple.com.

The Cloudflare page on screen is the Containers-backed Sandboxes product, which is a different runtime from the isolates in Cloudflare's own Code Mode post

At 59.36 s the video says "You are going to need some sort of sandbox environment", and at 62.36 s, "Cloudflare does have a code execution sandbox, but I will say you need to make sure you set this up right so you can't just have any agent executing these tools." Frames 031 through 034 show the Cloudflare Sandboxes product page, and the copy matches the live page verbatim: the headline "Secure code execution for AI agents and developer tools", the subhead about running untrusted code in isolated environments, and the three feature cards, "Fully isolated" ("Each sandbox runs in its own secure container. Execute AI-generated or user-submitted code with zero risk to your infrastructure."), "Fast spin-up" ("Sandboxes start in milliseconds, not minutes. Execute code immediately without waiting for cold starts.") and "Bring your own images" ("Use standard container images with your own dependencies, tools, and runtimes. Full flexibility without lock-in."). The product exists and the page is quoted accurately.

The runtime distinction is worth stating because the video treats the two as one thing. Frame 033 at 64 s shows the "Built on Containers" card explaining that Sandboxes run on Cloudflare Containers with global placement, automatic scaling and pay-per-use pricing, and frame 034 at 66 s shows "Code interpreter built-in", describing Python or JavaScript execution through runCode(). Cloudflare's Code Mode post, by contrast, runs the agent's generated code in V8 isolates on Workers, and argues specifically that "isolates are far more lightweight than containers", starting in milliseconds on minimal memory with a fresh isolate created per snippet and no reuse or prewarming. Both are Cloudflare code execution surfaces. They are different products with different isolation models and different cost profiles, and picking one over the other is an architecture decision the video does not surface.

The security caution at 62.36 s is well placed and matches Anthropic's own framing. Anthropic lists the tradeoffs of this pattern as a secure execution environment, sandboxing, resource limits, monitoring, and the operational overhead that comes with all of it, stating that "Running agent-generated code requires a secure execution environment with appropriate sandboxing, resource limits, and monitoring." The video reaches the same conclusion from a shorter argument.

The video discloses that nothing was tested, and the frames confirm it

At 71.36 s the speaker says "I haven't tested this out personally, but I'm planning on testing it," and at 73.96 s, "If you've already tested this out, let me know what your thoughts are." That disclosure is accurate and is worth crediting, because it is the part most videos in this format omit.

The frames corroborate it. Across the 19 frames I read, spanning 0 s to 78 s, the only non-talking-head content is two scroll captures, the AIMultiple article at 46 s to 50 s and the Cloudflare Sandboxes marketing page at 60 s to 66 s. There is no terminal, no editor, no code, no agent trace, no token counter and no before-and-after run. The grocery store analogy running from 16.62 s to 29.88 s is the only explanation of the mechanism offered, and it describes round-trip overhead without touching the part that actually drives the saving in both primary sources: tool definitions for every connected server being loaded into context upfront, which AIMultiple's own on-screen definition names as the distinguishing property of "Regular MCP".

Key Takeaways

  • Correction: The video says at 0.00 s that "traditional MCP is pretty much dead" and titles itself "Apparently MCP is obsolete for AI AGENTS". Cloudflare's Code Mode post states "MCP is still useful", Anthropic describes presenting "MCP servers as code APIs rather than direct tool calls", and the MCP specification shipped a GA revision, 2026-07-28, on July 28, 2026, with close to half a billion Tier 1 SDK downloads a month.
  • Verified: The 77% token reduction is real. AIMultiple's table reads Total All Tokens 775,197 for regular MCP against 175,081 for code execution, a -77.4% difference, and 175,081 divided by 775,197 confirms it.
  • Verified: The latency claim holds. AIMultiple reports 9.66 s average latency for regular MCP against 10.37 s for code execution, a +7% difference, and the video correctly says latency goes up.
  • Partial correction: The video calls the AIMultiple page "a bunch of different benchmarks" at 45.68 s. It is one benchmark: GPT-4.1, Bright Data's MCP Server in Pro Mode, two task types, 50 runs per approach, and AIMultiple itself lists limited query diversity and single-model testing as limitations.
  • Clarification: The on-screen table also shows Avg Output Tokens rising from 87 to 192 and Total Output Tokens from 4,345 to 9,585, both +120%, which the narration never mentions. At OpenAI's published gpt-4.1 prices of $2 per million input tokens and $8 per million output tokens, the benchmark's dollar saving is 74.1%, not 77.4%.
  • Correction: The video presents AIMultiple as the origin of the method. Cloudflare published Code Mode on September 26, 2025 and Anthropic published code execution with MCP on November 4, 2025. AIMultiple's article carries an updated date of September 9, 2026.
  • Context: Anthropic's own figure for the pattern is larger than the one the video cites: 150,000 tokens down to 2,000, a 98.7% reduction, though that is one illustrative Google Drive to Salesforce workflow rather than a multi-run benchmark.
  • Clarification: The Cloudflare page shown in frames 031 to 034 is the Containers-backed Sandboxes product. Cloudflare's own Code Mode post runs generated code in V8 isolates on Workers and argues isolates are lighter than containers. The video treats them as the same offering.
  • Verified: The speaker's security warning at 62.36 s matches Anthropic's stated tradeoffs, which name a secure execution environment, sandboxing, resource limits, monitoring and operational overhead.
  • Verified: The speaker discloses at 71.36 s that he has not tested the approach, and no frame I read shows a terminal, editor, code or any run of the technique.
  • Context: Whisper transcribes the publisher as "AI Multiple" in two words. Frames 024 through 026 show the logo as one word, "AIMultiple".
  • Unverified: TikTok's web page for this video could not be fetched, so the engagement counts of 1,207 views, 41 likes, 2 comments, 15 reposts and 33 saves come only from the local metadata capture at 2026-10-26T23:03:50Z and could not be confirmed against the live page.

Resources

Published October 26, 2026. Writeup generated from a favorited TikTok.