FIELDNOTES / COMMUNITY PROJECTS

What are people building?

Choose an action. Route a request. Check whether a job is finished. These projects show where a small decision fits into a larger workflow.

Sources reviewed September 20, 2026. Curated by task, not ranked by popularity. Results belong to the linked authors; we have not run these projects.

01 / DECISION TASKS

Act on a page

Browser Use / Jev Ultrafast

Choose the next browser action

Find flights from Zürich to London, stopping when matching options are visible.

What Jev decides: Jev picks an operation and an indexed page element. A separate text model writes values for text fields.
What you can inspect: The author reports a 7.073-second run, measured after the first page observation. The repository includes the recording, traces and measurement boundaries.

Code · video · author measurements

This is one flight-search task, not a completed booking or a general browser benchmark. We have not reproduced the run.

Read the project ↗Watch the demo ↗Inspect measurements ↗

Droidrun / Mobile Jev

Pick the next action on a phone

Navigate an Android ride-request flow.

What Jev decides: The agent turns device state into choices for the next action and target, then executes the selected action.
What you can inspect: The project publishes an Android recording. You can play it directly in our video collection.

Code · recorded demo

The recording stops at payment selection. It does not demonstrate a completed ride booking.

Play the video here →Read the project ↗

02 / DECISION TASKS

Route and supervise

LangChain / ModelRouterMiddleware

Choose a model for a request

Send a request to a fast model or a more capable model.

What Jev decides: Jev classifies the latest user message. The selected model is then used for every model call in that run.
What you can inspect: LangChain documents an experimental middleware implementation, including configuration examples.

Official integration · code example

The example explains the integration; it does not establish how much cost or latency your workload will save.

Read the integration ↗Read the harness article ↗

LangChain / AutoModeMiddleware

Check a tool call before execution

Inspect proposed tool calls before an agent executes them.

What Jev decides: Jev classifies calls to configured tools. Calls judged risky return an error ToolMessage instead of executing.
What you can inspect: The experimental middleware gives developers a concrete place to add a model-based check.

Official integration · code example

It does not request human approval. Add a separate approval flow when needed; a model judgment is not a safety guarantee.

Read AutoMode documentation ↗

Thruwire / Foreman

Check whether a coding agent is done

Supervise a Codex worker as it implements a task.

What Jev decides: Jev judges observations such as completion, requirement coverage and whether more verification is needed. Python logic decides how to continue, steer or stop the worker.
What you can inspect: The repository separates the worker from a concurrent observation loop, making the supervision approach inspectable.

Code · architecture experiment

The author describes an architecture experiment, with no claim that it outperforms a conventional harness.

Read the project ↗

GiesN / Jev email intent workflow

Send an email to the right handler

Separate invoice-related requests from general mail.

What Jev decides: A Jev Choice returns invoice or general. A LangGraph branch selects accounts payable or the general inbox.
What you can inspect: The example includes ten mocked emails, expected labels and routing tests. Running the classifier requires real API calls.

Code · ten labeled sample emails

Handlers only set a destination in graph state. They do not send mail or make payments; ten samples are a smoke check, not an accuracy benchmark.

Read the project ↗Build a routing example →

03 / DECISION TASKS

Score and filter

Valentynkit / jev-skip

Find sponsor segments in a video

Mark likely sponsor segments using a YouTube video’s captions.

What Jev decides: Jev classifies transcript windows. The extension uses category probabilities to paint the timeline and decide which segments to skip.
What you can inspect: The author reports catching 77% of SponsorBlock-labeled sponsor seconds across 23 videos, with 34 seconds of false skips per hour.

Code · recorded answers · fixture evaluation

Those results use recorded answers through a gateway shim. The demo replays answers at recorded latency; it is not a live API demonstration. No captions means no classification.

Read results and caveats ↗See the demo recording ↗

Stemonte / Kill My Idea

Score an idea against a rubric

Evaluate a submitted idea against several questions at once.

What Jev decides: The author describes roughly ten parallel questions or classifications, without a generative text model.
What you can inspect: The original post explains the side project and links to the author’s app.

Author’s community post

The post is a product demonstration, not a validation of the scores. A rubric score does not establish whether an idea will succeed.

Read the original post ↗

FOLLOW THE DISCUSSION

Original social posts to explore

LangChain’s harness article ↗ points to these developer demos. The X posts could not be fully inspected during our review, so they are reading leads rather than verified case studies.

Kyle Jeong / browser automation ↗Ryan Vogel / email triage ↗

Looking for more repositories? Browse awesome-jev ↗. Inclusion in a directory is not evidence of reliability.

Ready to try one small decision?

Start with a routing example, or watch the official Doom and Wikiracing demos.

Build a router →Watch demos →