Decision models go local while agent harnesses argue over retries, YAML and plugins

Fastest-growing this period

Counts come from GitHub's official stargazers/history endpoint: the stars gained on each day of the window. Project briefs and the reading notes below are editorial additions.

1. 1

   NandhaKishorM/laya : Answer typed questions about any text in one 33-millisecond forward pass

   Python

   A non-autoregressive decision engine that reads text, an email or a JSON document and returns a typed answer, a choice, a score or a yes/no flag, in one 33-millisecond forward pass. Three checkpoints cover English, over one hundred languages and typed workflows, and a built-in router picks one per request. Nothing is generated, so there is nothing to parse or hallucinate.

   +5,604
2. 2

   jev-chat/jev-chat-jarvis : A phone copilot that reads on-screen chats and drafts replies, sending stays yours

   Kotlin

   An Android assistant that follows conversations in any chat app via the accessibility service, never hooking or modifying the app. It first judges the other side's intent and a danger level, then drafts ranked candidate replies and pastes the one you pick into the input box; it never presses send and keys stay on the device. WeChat, QQ and X run the full pipeline.

   +2,640
3. 3

   jaredpalmer/kev : Small open decision models answering choice, yes/no and scoring questions in one request

   Python

   A family of small decision models built on Qwen3.5, with 0.8B, 4B and 9B variants you can run or train yourself. One request shares a single input across choice, yes/no and rating questions, each answered with probabilities instead of a single label. It serves locally on CUDA, ROCm and Apple Silicon, and a web playground shows how option order changes the answers.

   +2,063
4. 4

   browser-use/jev-ultrafast : A browser agent that picks each click and typed word in one model call

   Python

   Give it one natural-language goal and it renders the page as a numbered table of controls. TypeSafe's Jev picks the operation and target element in one request, and a small LLM writes text only when a field needs filling, so no site scripts or screenshots drive the loop. The README's measured demo finds flights from Zürich to London in about seven seconds.

   +1,734
5. 5

   hypit-ai/hypit : Let your coding agent turn one reference video into a hundred variants

   TypeScript

   A skill that gives coding agents a language for making video: drop in a reference clip and the agent rebuilds it as a complete workflow covering footage, captions, B-roll and effects, timed to words instead of seconds. You can also start from templates or a plain description, and a workflow can compile captions and motion graphics without calling a generation model.

   +1,664
6. 6

   google/ax : Declare agent tasks in YAML, sandbox them, then SSH into a running one

   Go

   A declarative orchestrator from Google for agent workloads on a cluster, in the spirit of Kubernetes. One YAML file declares the task, pre-wired Git repos and MCP servers, a network allowlist and the model; ax apply launches it into a sandbox with CPU and memory limits, and ax ssh opens a shell inside a running agent to watch what it does.

   +1,580
7. 7

   unreallabsai/unreal-agent : An async-first agent harness where redelivered inputs never run twice

   Go

   An async-first agent harness in Go where retries are safe. Every input carries a caller-supplied unique ID and is deduplicated in a session inbox, so redelivered events never run twice. Sessions are append-only histories that can be forked, tool calls become serializable operations executed asynchronously, and formats are versioned so old sessions resume or fail loudly.

   +1,255
8. 8

   mizorewww/laya-mlx : Typed decisions from the laya family, rebuilt to run natively on Apple Silicon

   Python

   A native Apple Silicon build of the laya typed-decision engine on Apple's MLX runtime, with no PyTorch, no Transformers dependency and no cloud API. A short English typed question is answered in about 13 milliseconds with zero output tokens, and the multilingual checkpoint is faster still. Weights download once, then every call runs fully local.

   +1,247
9. 9

   CopilotKit/openmuse : A self-hosted personal agent whose browser and terminal you can take over

   TypeScript

   A self-hosted personal agent app from CopilotKit, one codebase for iOS, Android and web. It runs its own server, persistent Chromium and an optional Linux container, moves PDFs and shows plans and approvals inline. Take control opens its browser or terminal so you can continue the work yourself; Gmail and calendar connect through Google OAuth with every send held for review.

   +1,192
10. 10

    deepseek-ai/deepseek-harness : Explore DeepSeek's building blocks for an extensible agent

    TypeScript

    DeepSeek's agent runtime built around replaceable plugins, with a web interface for running it. Useful for developers exploring how to compose an agent's capabilities; it is a developer preview, so expect interfaces to change as you experiment.

    +1,166

Ranking window: 2026-09-22 UTC, full day.