Last Week’s List Had Five Jev Projects. This Week’s Has None.

Last issue ended on a question: would the Jev wave have a second week? It did not. Five projects on that list were built on or around TypeSafe’s decision model — Jev driving a browser, an open copy of it, an open rival, the rival’s port to Apple silicon, and TypeSafe’s own skill pack — and this week none of them reached either board. Laya came nearest and still fell short of tenth place. Every one of them is still adding stars, at a fraction of the pace they launched with, which is what a launch week looks like when it is imitated very quickly: one burst, copied within days, cooling together. The board that replaced them has a different texture. Two of the newcomers rose on dated, checkable events — a model far too large for a gaming PC made to run on one, and NVIDIA’s new agent-safety platform. And most of what we said last issue about the regulars broke: Hindsight stopped accelerating almost the moment we pointed at it, VoiceStudio’s steady pace ended in a burst, and Orca, which we said had stopped wobbling, went straight back to wobbling.

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

+13,086 stars this week → 53,154 · VoiceStudio star curve

2026-W37: +7,528 · 2026-W38: +6,912 · 2026-W39: +6,788 · 2026-W40: +6,541 · up and down across 5 issues · Earlier: 18,960 stars on 2026-09-06

Last issue the line beside this project said its pace had not changed, and for four issues it had not: a voice suite adding nearly the same amount every week, as regular as anything on the board. The week we printed that was already ending in a break. Its daily readings turned sharply upward on the last day of last issue’s window, kept climbing for two more days, and then decayed back toward the old rhythm by the close — a burst bolted onto a plateau, not a new trend. Almost all of this week’s lead over its usual total came from those few days. We could not find a single event that explains the timing. The releases just before it were product work rather than news: a move to a packaged desktop app, cross-language cloning that speaks your voice in a language you do not, and hooks for Claude Code, Cursor and phone calls through Twilio. What did arrive in the same stretch was a run of write-ups calling it a fully local ElevenLabs alternative, the framing its own README leads with. That framing is the likelier engine: it turns a toolkit into a substitute for a paid service, and substitutes travel. The thing to check next issue is whether the floor it settled on afterwards, already a little higher than the one it left, holds.

#2 Niko1221/Strata — a local inference engine for running 125B parameter models on consumer gaming PCs

+10,436 stars this week → 10,966 · Strata star curve

Earlier: 530 stars on 2026-09-28

This is the only newcomer in the top ten that was still speeding up when the week closed: nearly every daily reading larger than the one before it, from a repository almost nobody had seen at the start. Strata is not a model. It is an installer and inference engine for one model — Alibaba’s Qwen3.8-Flash-Next, a mixture-of-experts design that wakes only a small slice of its experts for each token — and the trick is where it keeps the rest. Hot weights live on the graphics card, colder ones in system memory, and the remainder is paged in from the SSD as needed, which is how a model of that size fits beside a mid-range gaming card. The launch drew a long Hacker News thread, and the pushback there is the part worth keeping. The weights are compressed well below the usual four-bit line, testers reported image understanding degrading sharply at that level, the base model is itself labelled an under-trained preview, and the throughput figures in the README are the author’s own. None of that makes the demo less real. It does mean the open question is not whether a gaming PC can run it, but whether what it runs is worth asking anything hard. A curve still rising at the close is the shape to check first next issue — which is exactly what we said about Hindsight last week.

#3 vectorize-io/hindsight — an agent memory system that enables AI agents to learn and synthesize information over time

+8,258 stars this week → 45,490 · hindsight star curve

2026-W40: +13,143 · down from the last issue · Earlier: 23,009 stars on 2026-09-07

Last issue we called this the one project on the board still accelerating at the close, and the shape to check first. We checked. The acceleration ended almost as soon as we wrote it down: the largest day we have recorded for it landed on the boundary between the two windows, the next day was smaller, the one after smaller again, and by midweek it had settled onto a flat line it held to the end. The burst was real and the trend was not. It is now adding a modest, nearly constant amount each day — more than it managed at the start of September, nowhere near its peak — and it is worth saying plainly that reading one rising tail as momentum was premature. What we got right was the product. Vectorize’s late-summer notes describe a single command that wires memory into a long list of coding agents, all sharing one bank per repository, and its newest release keeps adding to that surface rather than to its headline benchmark, which no one outside the company has yet reproduced. A memory layer that every harness on this board can plug into grows differently from a launch: slowly, through other people’s tools. A flat line that stays flat may be what that looks like.

#4 DietrichGebert/ponytail — a framework to prevent AI agents from over-engineering code by enforcing a lean decision ladder

+7,933 stars this week → 154,851 · ponytail star curve

Earlier: 127,917 stars on 2026-09-06

#5 paperclipai/paperclip — an open-source orchestration control plane for managing autonomous teams of AI agents

+7,379 stars this week → 97,179 · paperclip star curve

2026-W40: +8,683 · down from the last issue · Earlier: 80,084 stars on 2026-09-06

#6 NVIDIA/OpenShell — a secure sandboxed runtime for autonomous AI agents with policy-enforced egress and credential injection

+6,053 stars this week → 14,862 · OpenShell star curve

Earlier: 8,520 stars on 2026-09-06

For most of September this repository barely moved — a handful of stars a day, the trickle of infrastructure code that people bookmark and leave. Then, on the last Monday of the month, NVIDIA announced its Open Agent Safety Platform, and OpenShell is the open half of it. It is a runtime that wraps an existing agent without rewriting it, confines it to a sandbox, decides which hosts and APIs it may reach, and keeps credentials outside the agent’s own process, so a compromised agent has nothing to read. The other half, Sentry, is a watchdog meant to run on NVIDIA’s network processors and quarantine an agent that steps outside its policy; it ships as a reference design, not as code anyone can star. The repository is older than the announcement — it first appeared inside NVIDIA’s reference agent stack in the spring — and its curve has the shape Google’s AX showed last issue: dormant code, a keynote, three days of climbing, then a fast cool-down to a small fraction of the peak. Read beside Paperclip and Orca on the same board, it fills the gap they leave. They decide which agent does what; this decides what any of them is allowed to touch. That is a less exciting job, and a launch-partner list of over a hundred companies suggests it is the one enterprises were waiting for.

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

+5,727 stars this week → 243,391 · deepseek-harness star curve

2026-W35: +42,534 · 2026-W36: +16,749 · 2026-W37: +8,162 · 2026-W38: +8,562 · 2026-W39: +8,642 · 2026-W40: +6,473 · up and down across 7 issues · Earlier: 213,124 stars on 2026-09-06

#8 mattpocock/skills — a set of composable agent skills that apply engineering fundamentals to AI-assisted coding

+5,478 stars this week → 276,162 · skills star curve

Earlier: 252,593 stars on 2026-09-06

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

+5,413 stars this week → 84,974 · orca star curve

2026-W35: +5,420 · 2026-W36: +5,522 · 2026-W37: +5,082 · 2026-W38: +5,022 · 2026-W39: +5,789 · 2026-W40: +5,983 · up and down across 7 issues · Earlier: 62,230 stars on 2026-09-06

Last issue we split the two names that have appeared in every issue. We said DeepSeek Harness had resumed its slide after a plateau, and that Orca had broken its old rate twice running and was rising rather than wobbling. This week the first call held and the second broke. DeepSeek Harness stepped down again — a modest step, not a drop — after a short midweek bump that faded by the close. Orca fell back too, below both of its record weeks, and with that the verdict printed above this paragraph is fair again: over its whole run it has moved around one level, and the two records look like the top of that wobble rather than a break out of it. Our guess last time was that Orca’s audience grows with everyone who starts running more than one agent at a time. This board offers a different reading. Orca now has company doing parts of its job: Paperclip coordinates whole teams of agents, OpenRig in the breakout list organises them into persistent groups inside tmux, and OpenShell fences what each of them may touch. A tool that once stood almost alone for running agents side by side is now one option among several, and a steady level is what a useful tool in a crowded category looks like.

#10 Panniantong/Agent-Reach — Agent Reach – One‑click Internet Access for AI Agents

+5,069 stars this week → 90,858 · Agent-Reach star curve

Earlier: 78,240 stars on 2026-09-06

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.

mvschwarz/openrig · +738.8% over 7.7 days → 4,974 stars · openrig star curve

Earlier: 593 stars on 2026-09-27

OpenRig is a multi-agent harness that organizes AI coding agents into persistent, coordinated teams using tmux and YAML-defined topologies, preventing terminal sprawl by managing agents as a single system.

CopilotKit/OpenDots · +518.2% over 2.8 days → 3,196 stars · OpenDots star curve

OpenDots is an open-source template for persistent AI coworkers with isolated compute and multi-channel communication, providing agents with their own browser profiles and terminals.

morluto/rea · +374.2% over 7.7 days → 1,968 stars · rea star curve

Earlier: 396 stars on 2026-09-06

REA is an agent-driven reverse engineering framework for analyzing app behavior and native binaries, bridging AI agents to professional tools like Hopper and Ghidra via MCP.

KKKKhazix/AIHOT · +374.2% over 5.8 days → 5,771 stars · AIHOT star curve

AIHOT is an LLM-powered, customizable news-aggregation and daily-briefing platform that filters and clusters news items from configurable sources into structured reports.

yetone/magpie · +288.1% this week → 4,813 stars · magpie star curve

Earlier: 673 stars on 2026-09-25

Magpie is a centralized model manager and local gateway for switching AI models across multiple local agents, letting one subscription serve several tools through a single interface.

The question for next issue is narrower than last week’s. Strata, Ponytail and Agent Reach all spent this week speeding up rather than slowing, and Strata is the only one of the three that is new. Last issue the single curve still rising at the close stopped rising the next day. If Strata does the same, this board is a sequence of launch weeks; if it holds, it is the first project since the Jev wave to earn a second week of its own.