<- all tokdocs
2026.10.23@agentic.conversations

The clip says an MCP agent "calls it through the session"; the current MCP specification says the protocol is stateless and an open connection is not a session

AI
Watch on TikTok

View on TikTok ->

A 41.64 second vertical clip, 720x1280 in an MP4 container (mov,mp4,m4a,3gp,3g2,mj2), video in HEVC Main profile level 120, yuv420p, 30/1 fps, 1,249 video frames, 555,926 bps, audio in AAC at HE-AACv2 profile, 44,100 Hz stereo, 64,306 bps, 901 audio frames, overall container bitrate 627,033 bps, filesize 3,263,868 bytes, confirmed by ffprobe against the downloaded video.mp4; yt-dlp recorded format id bytevc1_720p_627018-1, which matches the measured bitrate to within a thousandth. The extracted audio.mp3 runs 41.717551 s at 124,951 bps, 651,586 bytes. The post timestamp is epoch 1792771200, which is 2026-10-23T16:00:00Z, nine in the morning Pacific Daylight Time on the second and final day of the conference the description promotes. The uploader is agentic.conversations with nickname "Agentic Conversations" and uploader id 7535882886172460046; the audio credit is "original sound - Agentic Conversations." The description runs 366 characters and ends with five hashtags: #agntcon #webmcp #mcp #aiagents #webdev. At the capture time of metadata.json, 2026-10-24T06:05:10-0500, the counters read 391 views, 9 likes, 9 saves, 1 repost and 0 comments. I extracted frames at 0.5 fps into tokdoc_frames/ and read 20 of the 21: frames 001 through 013 and 015 through 021, covering 0 s to 40 s at two second steps. Because the caption text changes faster than two seconds in the opening, I re-extracted at 2 fps into tokdoc_frames_1s/ and read frames 012, 014, 015, 016 and 018, covering 5.5 s, 6.5 s, 7.0 s, 7.5 s and 8.5 s. The transcript is 110 words across 13 segments. The clip opens on a man with dark curly hair and a green, blue and pink vertically striped shirt with a lapel microphone on his collar, standing on an outdoor street; a black-on-white title card reads "The difference between WebMCP and MCP" (frame 001, 0 s). From 2 s to 7 s a full-screen architecture graphic replaces the footage: a pale blue box labelled "HOST" holding green boxes "MCP client 1," "MCP client 2," "MCP client 3" beside three host icons (a helmeted robot face, a grey cube, and the orange eight-point Anthropic starburst), wired to a second pale blue box labelled "MCP SERVERS" holding "MCP server 1," "MCP server 2," "MCP server 3," with one bracket across the top labelled "Transport layer"; captions over it read "So using MCP" (frame 002, 2 s), "doesn't need me" (frame 003, 4 s), "It's calling it" (2 fps frame 012, 5.5 s), "through" (frame 004, 6 s) and "the session" (2 fps frame 015, 7.0 s). The footage returns at 7.5 s to a woman in a cream blouse with a black neck bow and dark cat-eye glasses, captions "it probably" (2 fps frame 016, 7.5 s) and "to open a browser" (frame 005, 8 s). At 10 s a second full-screen graphic appears: a browser window titled "My Task Board" with an empty "Add a task..." field, an unchecked "Write blog outline" and checked "Research WebMCP" and "Draft with examples," linked by dashed arrows through a purple </> badge to a white robot whose speech bubble reads "I'll use available tools to get this done," headed "web MCP" with a large "WebMCP" wordmark beneath (frame 006, 10 s). From 12 s to 34 s the clip stays on the two speakers outdoors with word-by-word captions: "my logged in" (12 s), "is what the agent" (14 s), "Which also means" (16 s), "the tools" (18 s), "are contextual" (20 s), "every tool" (22 s), "It's like" (24 s), "on the cart page" (28 s), "tools that are" (30 s), "Oh I didn't" (32 s), "the page that" (34 s). From roughly 35 s a dark blue end card slides in and holds to 41.6 s: "AGNTCon + MCPCon North America," "The conference for AI agents & MCP," a circular headshot beside a pill badge "SPEAKER AT THE CONFERENCE" and the name "Angie Jones," a QR code, "SCAN FOR EVENT DETAILS" and "Oct 22-23 • San Jose" over a neon skyline (frames 019, 020, 021 at 36 s, 38 s, 40 s). No frame shows a convention centre, a badge, a lanyard, a terminal, an editor or a running WebMCP demo; both graphics are static illustrations, not screen recordings. The spoken argument is that an MCP agent calls a service without the user and without a browser, while WebMCP runs inside the user's already logged-in browser session and exposes only the tools that belong to the page the user is currently on. The last transcript segment ends at 36.540 s against a real audio duration of 41.642 s in the container and 41.718 s in the extracted MP3, so no segment is timestamped past the end of the audio; the 5.1 second tail is the silent end card, and nothing in this transcript is a Whisper hallucination past the end.

The event name, dates and city are right, and the post landed on the correct day

The description opens "Today is the last day of AgntCon + MCPCon North America in San Jose." Every part of that checks out against the organiser. The Linux Foundation press release dated April 2, 2026 lists "AGNTCon + MCPCon North America" on "October 22-23" in "San Jose, California," produced by the Agentic AI Foundation with The Linux Foundation. The Linux Foundation event schedule page for the conference confirms the same two dates and city. The post timestamp of epoch 1792771200 converts to 2026-10-23T16:00:00Z, which is 09:00 Pacific Daylight Time on October 23, the second of the two days. "Last day" is accurate.

The casing in the description is the only thing off, and it is cosmetic. The organiser writes the event as "AGNTCon + MCPCon," with AGNT in full capitals, which is also what the end card in frames 019 through 021 renders. The description writes "AgntCon."

The end card's "SPEAKER AT THE CONFERENCE: Angie Jones" is verifiable. The Linux Foundation schedule lists Angie Jones delivering "Welcome to AGNTCon + MCPCon North America," a keynote from 9:00 AM to 9:05 AM on Thursday, October 22, in the Grand Ballroom on the Concourse Level. The AGNTCon + MCPCon Japan 2026 schedule gives her title in full: "Angie Jones, Vice President of Developer Experience, Agentic AI Foundation." No frame in this clip identifies either on-camera speaker by name, so I will not assert that the woman in the cream blouse is Angie Jones. The end card names her as a conference speaker, which is a different claim from naming the person on camera.

"It's calling it through the session" does not survive contact with the current MCP specification

From 5.860 s to 7.560 s the speaker says "It's calling it through the session," and the burned-in caption confirms the wording across two frames: "It's calling it" at 5.5 s and "the session" at 7.0 s. This is not a Whisper mangle. The word is "session," and it is the wrong word for what MCP does today.

The MCP specification at revision 2026-07-28 has a section headed "Statelessness" that reads: "The Model Context Protocol (MCP) is a stateless protocol: all the information needed to process a request is contained in the request itself. A server processes each request independently; no state should be inferred from previous requests, even those on the same connection or stream." It then makes the point explicitly in a note: "This implies that an open connection, such as a STDIO process, is not a conversation or session: clients may interleave unrelated requests on the same transport, and a server must not treat connection or process identity as a proxy for conversation or session continuity." The normative bullets go further: "Servers MUST NOT rely on prior requests over the same connection to establish context," and "State that needs to span multiple requests ... MUST be referenced by an explicit identifier the client passes on each request."

The architecture overview for the same revision says the same thing in plainer language: "MCP is a stateless protocol (Every request contains all the information needed to process it, so servers infer nothing from previous requests). Every request carries the protocol version and the capabilities ... relevant to that request in its _meta field, so the server can process each request on its own."

The speaker is not inventing the word out of nothing. MCP did have a session concept, recently enough that the habit persists. The 2025-06-18 revision of the transports page has a section headed "Session Management" defining it: "An MCP 'session' consists of logically related interactions between a client and a server, beginning with the initialization phase." That revision specified an Mcp-Session-Id header on the Streamable HTTP transport, assigned at initialization, required on all subsequent requests, and terminable by the server with a 404. That is the model the speaker is describing, one revision behind the current one, and the direction of travel is away from it.

The practical correction is narrow and worth stating plainly. An MCP agent calling a remote service is not reusing a browser session and is not inheriting the user's logged-in state. Under the 2026-07-28 revision it is sending independent JSON-RPC requests, each carrying its own metadata, and authenticating per the Authorization framework for HTTP-based transports. Whatever credentials it uses had to be provisioned to the client separately. That gap is exactly the gap WebMCP exists to close, which the rest of the clip gets right.

The graphic playing over that sentence does not help it. The graphic occupying 2 s through 7 s resembles the MCP architecture picture but does not match the official one. The current architecture overview at modelcontextprotocol.io renders participants as a Mermaid graph with four clients, "MCP Client 1" through "MCP Client 4," inside a subgraph labelled "MCP Host (AI Application)," each wired by a "Dedicated connection" to a named server: "MCP Server A - Local (e.g. Filesystem)," "MCP Server B - Local (e.g. Database)," and "MCP Server C - Remote (e.g. Sentry)," with two clients sharing Server C. The video's version has three clients, three generically numbered servers, a box labelled "MCP SERVERS" that groups them, and a single bracket labelled "Transport layer" spanning the whole picture. It is a third-party redraw, not the source diagram.

The redraw is also not wrong, as far as it goes. The official overview defines the two layers the bracket gestures at, the "Data layer" that "Defines the JSON-RPC based protocol for client-server communication" and the "Transport layer" that "Defines the communication mechanisms and channels that enable data exchange between clients and servers, including transport-specific connection establishment, message framing, and authorization." It confirms the one-to-one pairing the arrows show: "The MCP host accomplishes this by creating one MCP client for each MCP server."

What the diagram does not contain is a session. There is no box, label or arrow on screen that corresponds to the word the speaker uses at the exact moment the graphic is on screen. The caption "the session" sits over a picture whose only connection-level label is "Transport layer," and the current specification says that an open connection on that transport layer "is not a conversation or session." The graphic and the narration are describing different things.

The WebMCP claims are correct, and the explainer says so in close to the same words

From 9.840 s to 16.700 s the speaker says "Versus WebMCP, then my session, my logged in session is what the agent now has access to," and from 16.820 s to 21.340 s, "which also means the tools that are available in that website are contextual." From 21.620 s she adds, "So it's not like, oh, here's every tool ever for my app," and at 29.260 s the exchange lands on "You're on the cart page. Yes. Here are the tools that are available to call."

Both halves of this check out against the WebMCP explainer maintained by the W3C Web Machine Learning Community Group. On the logged-in session, the explainer's stated motivation is avoiding the need to "replicate the user's state, active context, and authentication credentials on a separate server," and one of its goals is preventing the "disintermediation of web apps by backend integrations by adapting front-ends for use by agents." The agent operates inside the page the user already authenticated, which is what the speaker describes.

On page-contextual tools, the explainer's guidance is almost the sentence spoken in the clip: "Rather than registering a large catalog of tools upfront, dynamically register tools relevant to the active page state or workflow." The mechanism covers the whole lifecycle. Tools are registered with registerTool() carrying a name, description, inputSchema and an execute callback; they can be unregistered via an AbortSignal when they stop applying; and a toolchange event notifies listeners when the registered set changes. The cart page example is the explainer's design intent restated, not an exaggeration of it. The man's closing line at 32.300 s, "Oh, I didn't realize it was even specific to the page that you're on," is the correct reaction: page-scoped tool registration is the substantive difference between the two models.

The API neither speaker names has already moved, and the clip's framing of it is now a revision behind in a second way

The clip never names an API surface, so there is nothing here to contradict. It is worth recording anyway, because anyone acting on this 41 second explainer will reach for the wrong spelling.

Chrome's WebMCP documentation, published May 18, 2026 and last updated October 7, 2026, states that the feature is available "from Chrome 149" through an origin trial, and that local testing is enabled via the flag chrome://flags/#enable-webmcp-testing. It describes two registration paths, an "Imperative API" to "Define different types of tools with standard JavaScript, such as form input, navigation tools, state management, or other functions," and a "Declarative API" to "Add annotations to a standard HTML forms to create a WebMCP tool." It also describes the page-context behaviour the clip is about: WebMCP supports "a shared understanding of the current page context, so the agent knows what resources are available to act on in real time." Tool registration is gated by the tools Permissions Policy, which "defaults to self, allowing tool registration in top-level and same-origin contexts."

The surface itself changed. The explainer in the webmachinelearning/webmcp repository now documents document.modelContext, not navigator.modelContext. The Chrome documentation page does not discuss the move, and I could not reach a first-party Chrome or Chrome Status page stating the deprecation milestone, so I record the commonly cited Chromium 150 figure as unverified rather than citing it. What I can state from the primary explainer is that document.modelContext is the documented surface and navigator.modelContext is not.

One further status point, because WebMCP is routinely described as a standard. The specification is a Draft Community Group Report from the W3C Web Machine Learning Community Group, which is neither a W3C Recommendation nor on the W3C Standards Track. The clip does not claim otherwise, and that restraint is to its credit, but the distinction matters for anyone deciding whether to build on it.

The transcript mis-hears two words that the burned-in captions get right, and may mis-hear the opening question

This corpus has a known failure mode where Whisper converts product names into plausible English. That did not happen here. "MCP" and "WebMCP" come through intact in all 13 segments, and the burned-in captions corroborate them. Two smaller errors did occur, and the frames caught both.

At 7.720 s the transcript reads "We probably doesn't even need to open a browser." The burned-in caption at 7.5 s reads "it probably," not "we probably." The sentence is "It probably doesn't even need to open a browser," which also fixes the subject-verb disagreement the transcript introduces. At 2.000 s the transcript reads "Yeah, using MCP, my agent doesn't need me," while the burned-in caption at 2 s reads "So using MCP." The caption is an edit rather than a strict transcription here, so I treat "Yeah" as audible filler the editor dropped; either way the clause the claim rests on is identical.

The opening line is less settled. Whisper renders the 0.000 s to 2.000 s segment as "What's wrong with MCP versus WebMCP?" The title card burned over that same moment reads "The difference between WebMCP and MCP," and the creator's own description frames the clip as a comparison, not a complaint: "With MCP, the agent calls your service directly ... With WebMCP, it works inside your logged-in session." Nothing in the remaining 39 seconds argues that anything is wrong with MCP. A plausible reading is that the spoken question was closer to "What's the difference between MCP versus WebMCP," and Whisper collapsed it. I cannot settle this from the artefacts I have, because the title card is a graphic overlay rather than a speech caption and no speech caption covers the first two seconds. I am flagging the opener as uncertain rather than quoting it as the clip's thesis.

No frame places this conversation in San Jose, and the organiser ran a Tokyo event five weeks earlier

The description says the conference is in San Jose, and it was. The footage is a separate question. Across all 20 frames of the 0.5 fps set, the setting is an outdoor residential street: narrow roadway, a car under a fitted fabric cover (frames 010, 012), tiled and panelled façades with green glass-block windows (frames 007, 009, 011, 015), potted plants and palms on the kerb (frames 010, 011, 016), a wall-mounted meter cabinet (frames 007, 015) and dense summer foliage (frames 001, 013, 017, 018). There is no convention centre, no signage in English, no badge or lanyard on either speaker, and no visible conference branding anywhere except the end card graphic. The architecture reads as Japanese residential rather than Californian.

I cannot prove where this was shot and will not claim it. What I can establish is that the organiser ran an event in Tokyo in the same programme year, and that the person named on this clip's end card spoke there. The April 2, 2026 Agentic AI Foundation announcement lists ten events, including "MCP Dev Summit Tokyo" on "September 10-11, Tokyo." By the time the schedule went live that event carried the conference branding: the AGNTCon + MCPCon Japan 2026 schedule site gives "2026 September 10-11 | Tokyo, Japan" and lists "Keynote: Welcome - Angie Jones, Vice President of Developer Experience, Agentic AI Foundation" on the Thursday at 09:00 JST.

The description is accurate about what was happening on October 23, and the end card is accurate about the San Jose dates. The footage is undated and unplaced by anything visible in it. Treating the conversation as something filmed on the San Jose show floor would be an inference the frames do not support.

Key Takeaways

  • Verified: AGNTCon + MCPCon North America ran October 22-23, 2026 in San Jose, California, organised by the Agentic AI Foundation with The Linux Foundation. The post's epoch timestamp 1792771200 is 2026-10-23T16:00:00Z, so "today is the last day" was true when it went up.
  • Correction: At 5.860 s the clip says an MCP agent is "calling it through the session." The MCP specification at revision 2026-07-28 states "The Model Context Protocol (MCP) is a stateless protocol" and that an open connection "is not a conversation or session." The session model the speaker describes, with its Mcp-Session-Id header, belongs to the superseded 2025-06-18 revision of the Streamable HTTP transport.
  • Verified: The WebMCP claims about page-scoped tools are accurate. The W3C Web Machine Learning Community Group explainer instructs developers to "dynamically register tools relevant to the active page state or workflow" rather than "registering a large catalog of tools upfront," with AbortSignal unregistration and a toolchange event.
  • Verified: The claim that WebMCP uses the user's logged-in session is accurate. The explainer's stated motivation is avoiding the need to "replicate the user's state, active context, and authentication credentials on a separate server."
  • Correction: The transcript's "We probably doesn't even need to open a browser" at 7.720 s is a mishearing. The burned-in caption at 7.5 s reads "it probably." The sentence is "It probably doesn't even need to open a browser."
  • Partial correction: The transcript's opening line, "What's wrong with MCP versus WebMCP?", conflicts with the title card over the same moment, which reads "The difference between WebMCP and MCP," and with the creator's own description, which frames the clip as a comparison. No speech caption covers the first two seconds, so the opener is unresolved. Do not quote it as the clip's thesis.
  • Clarification: The description writes "AgntCon." The organiser and the clip's own end card write "AGNTCon + MCPCon."
  • Clarification: The architecture graphic at 2 s to 7 s is a third-party redraw, not the diagram published at modelcontextprotocol.io, which shows four clients and three named servers rather than three and three. Nothing in the graphic depicts a session.
  • Context: The API the clip never names has moved. The webmachinelearning/webmcp explainer now documents document.modelContext, not navigator.modelContext. Chrome's documentation says WebMCP is available "from Chrome 149" via origin trial, with local testing behind chrome://flags/#enable-webmcp-testing.
  • Unverified: Secondary sources place the Navigator-to-Document move on May 27, 2026 and the deprecation of navigator.modelContext at Chromium 150. I could not reach a first-party Chrome or Chrome Status page stating that version, so I am not citing it.
  • Context: WebMCP is a Draft Community Group Report of the W3C Web Machine Learning Community Group. It is not a W3C Recommendation and is not on the W3C Standards Track. The clip does not claim otherwise.
  • Clarification: No frame shows San Jose, a convention centre, a badge or any conference signage apart from the end card graphic. The footage is an outdoor residential street with Japanese-style architecture. The same organiser ran an event in Tokyo on September 10-11, 2026 where the person named on the end card keynoted, but I cannot place this footage and do not claim to.
  • Context: Neither graphic is a screen recording. No frame shows a terminal, an editor, registered tools or a WebMCP demo running. The clip describes the mechanism without showing it.
  • Verified: The transcript is clean of tail hallucination. The last of 13 segments ends at 36.540 s against an audio duration of 41.642 s in the container, and the 5.1 second remainder is the silent end card.

Resources

Published October 23, 2026. Writeup generated from a favorited TikTok.