<- all tokdocs
2026.09.20@vibewithkevin

The Routing Layer Is the Cheapest Part of This Stack, Because Jev Is a Classifier Rather Than a Chat Model

AI
Watch on TikTok

View on TikTok ->

Every component in this build is real and publicly available, including Jev, which TypeSafe AI released on September 15, 2026, five days before this video went up. Kevin (@vibewithkevin) spends 83 seconds describing a voice-capture pipeline he has been running for about a week: he presses a button on a Pebble Index 01 ring, speaks, and the note ends up dispatched to either Claude Code or Codex without him touching a keyboard. The claim is that a decision model named Jev sits in the middle, asks five typed questions about each note, and uses the answers to pick the model and the effort level. He describes the whole architecture in speech, with animated title cards for the five questions and the two-branch dispatch. He shows one screen: the Pebble app's "Add HTTP MCP Server" dialog, with SSE selected and the group set to "Default MCP Sandbox."

Two transcription notes. Whisper wrote "Cloud Code" and "cloud code" throughout, and both are Claude Code. Whisper also rendered "the Pebble GitHub repo," which a frame in the video identifies as coredevices/mobileapp, with the two highlighted README rows being "Pebble Index 01 (ring) support: continuous BLE scanning, voice-note recording and ingestion" and "MCP (Model Context Protocol) client/tool integration."

The Full Path From Button Press to Dispatched Agent

The video says "you stand up your own MCP service and point the ring at it." The ring does not point at anything. The Index 01 is a stainless steel band with a microphone and a button, no screen, no sensors, and a non-rechargeable silver-oxide cell rated for roughly two years. It holds a Bluetooth LE link to the phone and nothing else. Every piece of software in this pipeline runs on the phone or past it.

  Ring (Pebble Index 01)          button held down, mic captures audio
        |  Bluetooth LE
        v
  Phone app (coredevices/mobileapp, Kotlin/Compose Multiplatform, GPLv3)
        |-- experimental module     BLE scanning, voice-note ingestion
        |-- index-ai module         on-device speech-to-text + classification
        |-- mcp module              MCP client
        |  HTTPS, SSE per the dialog shown in frame 16
        v
  Remote MCP server on a Cloudflare Worker  (self-hosted by Kevin)
        |
        v
  Jev                              5 typed questions, 1 call per note
        |
   +----+----------------------+
   v                           v
  Claude Code                 Codex
  design, writing             logic, infrastructure

Two things about this path are worth naming. The transcription happens on the phone before the note ever leaves it, which means the MCP server receives text and not audio. And the MCP client lives in the app, so what is actually configurable is an HTTP endpoint in a settings screen, which is why the demo screen is a form with a URL field rather than anything on the ring.

Where Jev runs is the one link the video does not specify. Jev is an API service, so the Worker would call it over HTTPS. Whether that call originates in the Worker or somewhere downstream is not stated.

The Five Questions and What Each One Decides

The title cards give the exact wording. Mapping each question to the decision it drives:

# Question as shown What it selects
1 Is this software work? Whether an agent runs at all, or the note goes to Obsidian as a plain note
2 Can it build it without me? Autonomous dispatch versus queueing for review
3 How hard is it? Effort or reasoning depth for the chosen agent
4 Which repo, or where in Obsidian? Destination: working directory for an agent, or vault path for a note
5 Design, infrastructure, logic or writing? Claude Code versus Codex

The shape here matches what Jev is built for. TypeSafe's product is a model that takes data plus a list of typed questions and returns one typed answer per question, each with a calibrated probability, rather than returning prose. Question 1 is a boolean. Questions 2 and 5 are picks from a fixed option set. Question 3 is a position on a scale. Question 4 is a pick from an enumerated list of repos and vault paths, which means that list has to be maintained by hand and a new repo is invisible to the router until someone adds it.

Question 2 is doing more work than the other four. It is the gate between "an agent writes code now" and "a human looks at this first," and a false positive there is the only answer in the set that causes an agent to act on a misunderstood instruction.

Jev Is a Classifier, and That Changes the Cost Objection

The obvious objection to putting a model in front of every voice note is that you have added an inference call to the critical path of a system whose entire pitch is speed. For an LLM-based router that objection holds. For Jev it mostly does not.

TypeSafe's published pricing is $0.042 per million input tokens with output priced at zero, and its own launch benchmark reports a decision completing in 0.114 seconds against 8.566 seconds for an LLM on the same workflow. MarkTechPost's coverage puts the cost per decision around $0.0004. Those are vendor figures and vendor-supplied benchmarks, and nobody has published independent numbers on them yet. Taken at face value, though, a fraction of a cent and a fraction of a second is a rounding error next to a Claude Code or Codex session that runs for minutes and costs orders of magnitude more.

The interesting cost in this design is the one the routing decision commits you to. A misroute on question 5 sends a writing task to Codex, which wastes one session. A misroute on question 2 starts an autonomous agent on work that needed a human first, and the cost of that is however far the agent gets before someone notices.

Splitting Claude Code and Codex by Work Type Is a Configuration Decision

The stated reason for the split is worth quoting precisely. Design and writing go to Claude Code "because that's where most of my skills live." That is a claim about his own accumulated configuration, not a claim about which model writes better prose. The on-screen card underlines it: the Claude Code branch is labeled "WHERE MY SKILLS LIVE" and the Codex branch is labeled "DIVISION OF LABOR."

Read that way, the split holds up fine. If your writing and design tooling is installed in one CLI, routing writing work there is correct regardless of the underlying model, and the honest framing is that the router is selecting an environment rather than a capability. Kevin says as much when he calls it a division of labor "for the moment."

The split also assumes both agents are reachable and authenticated the same way from wherever the dispatch happens. Codex is an open-source terminal agent that runs locally against your code, and Claude Code works the same way. Neither is a hosted endpoint you fire a request at. Something has to be listening on a machine with the repos checked out, which the video does not describe, and which is the part of this build that would be hardest to reproduce from the video alone.

What the Video Leaves Out

Five gaps, in rough order of how much they matter if you tried to copy this.

The dispatch mechanism. A remote MCP server on a Cloudflare Worker is straightforward, and Cloudflare publishes a template that deploys one without authentication at all. Getting from that Worker to a Claude Code session on a machine that has your repos is a separate problem, and it is the one the video skips entirely.

Authentication. The frame shows the app's MCP dialog with the group set to "Default MCP Sandbox," which suggests the app applies some sandboxing to MCP connections, and Cloudflare's guide covers adding OAuth through GitHub or Cloudflare Access. Whether this particular Worker requires a token is not shown. An unauthenticated Worker that can start coding agents against your repos is a URL you would not want guessed.

Confirmation. Question 2 decides whether an agent proceeds without him, and there is no mention of a notification, an approval step, or a way to cancel a dispatch after the fact. A voice note is a low-friction input, and low-friction inputs produce half-formed thoughts.

Misroute handling. The claim is that early testing "nailed every single voice memo or note to the right part of my AI brain or pulled up the correct repo." That is a personal impression over roughly a week of use, with no count of notes, no labeled set, and no record of near-misses. Jev returns calibrated probabilities, which means a confidence threshold and a fallback path are available in the API. Whether he uses one is not stated.

End-to-end latency. The Jev call is fast by the vendor's numbers, but on-device transcription, a BLE hop, and an HTTPS round trip to a Worker all sit in front of it. The video gives no wall-clock figure for press-to-dispatch.

Key Takeaways

  • The Pebble Index 01 is a $75 microphone and button on a ring, with no screen, no sensors, and a battery rated for about two years. It ships transcription to the phone, so the MCP server sees text.
  • The app doing the MCP work is coredevices/mobileapp, dual-licensed under GPLv3 and a Core Devices commercial license, with a dedicated mcp module and an experimental module for ring support. Core Devices also publishes the ring's hardware files.
  • Jev is real, recent, and exactly the shape this use case needs: typed questions in, typed answers with calibrated probabilities out, no prose. TypeSafe AI launched it on September 15, 2026 alongside a $40M seed round.
  • Putting a decision model in front of every voice note costs far less than it would with an LLM router. TypeSafe's own figures put a decision at roughly $0.0004 and around a tenth of a second.
  • The Claude Code and Codex split is described as following where his skills are installed, which makes it a statement about his configuration rather than a comparison of the two models.
  • Hosting a remote MCP server on a Cloudflare Worker is documented and supported, including a template that deploys without authentication. Adding auth is a separate step.
  • Unverified: the specific Jev question schema, the Worker code, and the claim that early testing routed every note correctly. That last one is an impression from about a week of personal use with no counts and no labeled evaluation set, so treat it as an anecdote. The performance and pricing numbers for Jev are also all vendor-published, with no independent benchmarks yet. The video never shows how a dispatch reaches a running Claude Code or Codex session, so that link is undescribed rather than disproven.

Resources

Published September 20, 2026. Writeup generated from a favorited TikTok.