Jev Had Two Projects on Last Week’s List. This Week It Has Five.

Last issue, two projects on this list were built on Jev, the decision model TypeSafe announced a fortnight ago. This issue it is five: Jev itself driving a browser, an open copy of it, an open rival, a port of that rival to Apple silicon, and TypeSafe’s own skill pack for coding agents — with a sixth, a chat co-pilot, that only borrows the name. The copies came fast. Kev and Laya were both public within three days of the announcement, and the Apple-silicon port of Laya followed the day after Laya itself. What the star curves add is that the wave is already slowing: every one of them took most of its week in the first few days and was adding a trickle by the end. Around it, the two names that have appeared in every issue split apart — one call we made last time held, and the other broke.

#1 vectorize-io/hindsight — Hindsight – a high‑accuracy, LLM‑agnostic memory layer for smarter autonomous agents

+13,143 stars this week → 37,232 · hindsight star curve

Earlier: 21,777 stars on 2026-08-30

For most of the month this was a quiet project: a memory service adding a steady trickle, never close to this list. Then, in the last few days of the window, it went vertical, and unlike almost everything else on the board it was still accelerating when the week closed — each reading larger than the one before it, the last the largest of all. That is the opposite of the Jev curves further down, which spent their week slowing, and it is the shape to check first next issue. What changed is not visible in the stars. Vectorize’s own notes on its late-summer releases describe a single install that wires a long list of coding agents into one shared memory bank per repository, knowledge pages that turn a bank into a wiki reachable over MCP, and images kept as memories a fact can cite. The first of those is the likeliest match for this board, which is otherwise crowded with harnesses and orchestrators: every one of them runs agents that forget between sessions, and this is the piece that remembers across all of them. The benchmark claims in its README are its own; we have not seen them reproduced by anyone else.

#2 google/ax — a high-throughput declarative orchestrator for running autonomous agent workloads in sandboxed clusters

+9,190 stars this week → 12,161 · ax star curve

Earlier: 1,973 stars on 2026-08-30

Google published this repository weeks ago and almost nobody noticed. It sat nearly motionless until the company presented it during its I/O live stream in mid-September, and then grew several times over in a matter of days. AX is not a model and does not ship one. It is a Kubernetes-style control plane written in Go, built around four primitives — tasks, workspaces, gateways and models — that isolate untrusted agent code, pre-wire Git repositories and MCP servers, fence outbound network traffic, and point at whatever inference endpoint you configure. The line the launch led with, that it can run billions of agents, is the line Hacker News pushed back on, and nothing in the repository lets you check it; a star count cannot vouch for it either. What the curve does say is narrower: this attention was an event, not a discovery. Like the Jev projects below it, AX took most of its week early and was adding a fraction as much by the end. Next to Paperclip and Orca, which also coordinate many agents at once, AX is the one written for whoever operates the cluster rather than the developer sitting at the terminal.

#3 browser-use/jev-ultrafast — Jev Ultrafast – Fast, LLM‑driven browser automation with a dynamic indexed action space

+9,019 stars this week → 20,808 · jev-ultrafast star curve

2026-W39: +8,617 · much the same as last week · Earlier: 3,172 stars on 2026-09-18

Last issue we called this the browser half of the Jev arrival and set it against Tencent’s bridge, which borrows the browser you are already signed into. Its README is more precise than the summary above this paragraph, and it corrects it: the language model is not what picks the action. TypeSafe’s Jev reads a numbered table of every control on the page and answers two typed questions — which operation, and on which element — while a small text model is called only when the chosen operation is typing. No screenshots enter the default loop. That division of labour is the whole thesis of the Jev wave in one repository: let a decision model make the decisions, and keep generation for the rare step that genuinely needs words. It is also why the rest of this issue exists, because Kev and Laya are attempts to own that decision model rather than rent it. The curve is the warning. Its week matches last week in total, but nearly all of it arrived in the first two days, and by the close it was adding a small fraction per day of what it added at the start. It is still a waitlist product with open code attached.

#4 paperclipai/paperclip — an open-source control plane for orchestrating teams of AI agents with org charts and budget controls

+8,683 stars this week → 89,800 · paperclip star curve

Earlier: 79,714 stars on 2026-08-31

#5 rocketride-org/rocketride-server — an open-source AI development environment for visually composing and deploying high-performance AI pipelines

+7,354 stars this week → 15,830 · rocketride-server star curve

Earlier: 7,374 stars on 2026-08-30

#6 debpalash/VoiceStudio — an open-source suite for voice cloning, design, and video dubbing across 646 languages

+6,541 stars this week → 40,068 · VoiceStudio star curve

2026-W37: +7,528 · 2026-W38: +6,912 · 2026-W39: +6,788 · holding the same pace, 4 issues running · Earlier: 11,977 stars on 2026-08-30

#7 deepseek-ai/deepseek-harness — DeepSeek Harness – a modular, plugin‑driven platform for building AI agents

+6,473 stars this week → 237,664 · deepseek-harness star curve

2026-W35: +42,534 · 2026-W36: +16,749 · 2026-W37: +8,162 · 2026-W38: +8,562 · 2026-W39: +8,642 · up and down across 6 issues · Earlier: 203,394 stars on 2026-08-30

Last issue we said DeepSeek Harness had landed slightly ahead of itself twice running, and read that as the end of a fall rather than a pause in one. This week says otherwise. Its gain dropped back hard, to the smallest week we have recorded for it — below even the trough it had climbed out of. The slide that began after its enormous debut in late August has resumed, and our reading of the two weeks in between was wrong: they were a plateau, not a floor. The contrast is Orca, the other name that has appeared in every issue. We said Orca had broken its fixed rate by arriving ahead of any week on record; it has now done that again, a second record in a row, which makes the verdict printed beside it misleading — its run has been steady and then rising, never really up and down. Nothing in either codebase explains the split. DeepSeek Harness is a general framework — a plugin system, a local web UI, hot-reloading tools — that drew its first crowd on the strength of the name on it. Orca is a narrow tool for running several coding agents side by side in separate worktrees. Our guess, and it is only a guess, is that one audience is shrinking back to the people who build frameworks, while the other grows with everyone who has started running more than one agent at a time.

#8 jaredpalmer/kev — a decision model that outputs calibrated probabilities for multiple typed questions in a single prefill pass

+6,402 stars this week → 7,413 · kev star curve

Earlier: 573 stars on 2026-09-20

Two days after TypeSafe announced Jev as a closed, paid API, Jared Palmer published an open imitation of it. Kev is a family of small decision models fine-tuned from Qwen with LoRA and a small pointer head, serving the same request shape as TypeSafe’s endpoint so that code written for one can be pointed at the other. It has grown quickly since: the first release ran on a laptop, and later ones scale up to a model that wants a data-centre GPU. The honest part is the comparison Palmer published himself. On data neither model was trained on, his largest checkpoint scores several points below Jev, and at a tight error budget it can automate a noticeably smaller share of decisions. That is the real trade the Jev wave is offering. Jev is quicker to adopt and more accurate; Kev is yours — you can train it on your own labels and run it where your data already lives. Its curve has the same shape as the others in this issue: most of the week in the first few days, a trickle at the end. Whether it climbs again depends less on attention than on whether someone publishes a fine-tune that beats the closed model on one narrow task, which is the outcome an open copy exists to produce.

#9 stablyai/orca — an AI orchestrator for running multiple CLI agents in parallel isolated worktrees

+5,983 stars this week → 79,561 · orca star curve

2026-W35: +5,420 · 2026-W36: +5,522 · 2026-W37: +5,082 · 2026-W38: +5,022 · 2026-W39: +5,789 · up and down across 6 issues · Earlier: 56,791 stars on 2026-08-30

#10 NandhaKishorM/laya — Laya – Multilingual, non‑autoregressive decision engine for instant, structured answers

+5,750 stars over 3.8 days → 26,698 · laya star curve

Laya is the second open answer to Jev, and it came a day after Kev from the opposite direction. Kev fine-tunes a generative model and bolts a decision head onto it; Convai Innovations built Laya on encoders from the BERT family, an architecture that predates chat models and never generated text in the first place. A router detects the language of the input and hands the question to an English checkpoint, a multilingual one covering more than a hundred languages, or one tuned for typed decisions, and each returns choices, scores or yes-or-no probabilities in a single pass, in tens of milliseconds on an old data-centre GPU. The weights are Apache-licensed. The speed follows from the architecture; the accuracy is a separate question, and we have not found it measured against Jev on the same data by anyone outside the project. Laya also has the fastest follower on either board this week: a port to Apple silicon through MLX appeared the next day and sits in the breakout list below, alongside TypeSafe’s own skills. Its short window shows the same deceleration as the rest of the lineage, only gentler — every reading smaller than the one before, none of them a collapse.

Breaking out

Ranked by growth relative to their own size, not by stars added. These are small enough that a week can multiply them — and small enough to have escaped you.

newliver666/apk-reverse · +449.8% this week → 2,705 stars · apk-reverse star curve

Earlier: 492 stars on 2026-09-21

jev-chat/jev-chat-jarvis · +327.1% over 5.8 days → 6,774 stars · jev-chat-jarvis star curve

mizorewww/laya-mlx · +239.0% this week → 6,482 stars · laya-mlx star curve

Earlier: 1,912 stars on 2026-09-21

typesafe-ai/skills · +179.6% over 7.7 days → 2,284 stars · skills star curve

Earlier: 817 stars on 2026-09-20

allenpeng0705/EnvoyMesh · +152.6% this week → 2,817 stars · EnvoyMesh star curve

Earlier: 576 stars on 2026-09-04

The question for next issue is whether the Jev wave has a second week. The original, the copy, the rival and the rival’s port all spent this one decelerating, and the only entry in the top ten still accelerating at the close — Hindsight — has nothing to do with any of them. If the Jev names hold their places, this is a category. If they fall away together, it was a launch week that happened to be imitated very quickly.