kas

KO | EN

gitlite — search

#javascript #python #ai #rust #deep-learning #linux #nodejs #java #awesome #vue #awesome-list #pytorch #openai #open-source

(★ 9)

Yet another local agentic shell

kas 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

**K.A.S — Kasra's Agentic Shell.** Run open models **locally** — on Apple Silicon
(MLX) or NVIDIA (llama.cpp/GGUF) — behind an **Anthropic Messages-compatible
server**, driven by an agentic TUI. Offline, no telemetry: tool use, streaming,
thinking, subagents, KV-cache continuation, and local recall, all on your iron.

## Install

```sh
curl -fsSL https://raw.githubusercontent.com/quantumwake/kas/main/install.sh | sh
```

Installs `kas` + `kas-server` as a [uv](https://docs.astral.sh/uv/) tool. It's
**GPU-aware**: MLX on Apple Silicon, a CUDA llama.cpp build on NVIDIA (prebuilt
wheel, ~10 s). Optional features (voice, vision, image-gen, web) install via
`kas doctor --install`. Uninstall: `…/main/uninstall.sh | sh`.

## Quick start

```sh
# 1. start the server — the FIRST run downloads the model (progress is shown)
kas serve --model mlx-community/Qwen3.6-27B-4bit          # Apple Silicon (MLX)
#   …or on NVIDIA:   kas serve --model unsloth/Qwen3.6-27B-MTP-GGUF

kas                # 2. launch the agent TUI

kas --yolo "build me an asteroids game in ./game, then run it"   # one-shot
```

A 27B 4-bit model is ~15 GB on first download. Once you have models, bare
`kas serve` loads the default and `kas` offers to start one and pick from what's
downloaded. Manage the daemon with `kas serve --status | --stop | --logs`. The
agent alone is portable — point it anywhere with `kas --base-url http://host:port`.

## NVIDIA / GGUF

Same agent, GGUF models through llama.cpp. Point `--model` at any Hugging Face
GGUF repo:

```sh
kas serve --model unsloth/Qwen3.6-27B-MTP-GGUF             # quant auto-picked
kas serve --model unsloth/gemma-4-31B-it-qat-GGUF --quant Q4_K_M
```

kas handles the GGUF rough edges for you: quant selection (`--quant` to pin),
context **auto-sized to the model + GPU** (a hybrid 27B fits 128 k on a 40 GB
card; a dense 31B backs off to fit), Flash Attention on, and live GPU memory in
`/stats`. Details in [`docs/backends.md`](docs/backends.md).

## Models

- 🍎 **MLX** — default `mlx-community/Qwen3.6-27B-4bit`; switch live with `/model`.
- 🟩 **GGUF** — any HF GGUF repo via `--model` (NVIDIA/CPU).

Tool calls are parsed per family (qwen · llama · mistral · harmony · hermes ·
deepseek · kimi · gemma), auto-detected from the chat template.

## Flags & config

```text
kas [task]                 interactive TUI, or one-shot if a task is given
  --yolo                   run bash without per-command confirmation
  --model ID --quant Q     pick a model (GGUF quant optional)
  --base-url URL           talk to a remote server (default 127.0.0.1:8765)
  --net / --art            enable web tools / image generation (opt-in)
  --resume [ID]            continue a saved session (warm KV)
kas serve --stop|--status|--logs|--no-daemon
```

Env vars mirror the flags (`KAS_MODEL`, `KAS_BASE_URL`, `KAS_MAX_TOKENS`, …).
GGUF tuning: `KAS_GGUF_QUANT`, `KAS_CTX` / `KAS_CTX_MAX`, `KAS_GPU_LAYERS`,
`KAS_FLASH_ATTN`, `KAS_BACKEND`. Run `kas doctor` for a capability report.

## What makes it tick

- **KV continuation** — agent turns prefill only the new tokens, not the whole
  transcript (≈ constant per-turn cost); persisted, so `--resume` rehydrates warm.
- **Per-thread caches + subagents** — the agent and each subagent get their own
  KV slot; delegate a subtask to a fresh context, only the report returns.
- **Compaction + recall** — compaction triggers on falling decode tok/s; local
  BM25 over code/docs/memory (`--rag`, offline) keeps it lossless.
- **Live + cancellable** — steer mid-task, `Esc` cancels even an in-flight
  prefill, `/model` hot-swaps the served model with no restart.

Architecture (hexagonal ports & adapters): [`docs/architecture/`](docs/architecture/).

## Tested

✅ run end-to-end · 🟡 runs, lightly checked

| GPU | Backend | Models |
|---|---|---|
| 🍎 Apple Silicon | MLX | ✅ Qwen3.6-27B · gpt-oss-20b · Gemma 3/4 · Llama-3.2 · Mistral-7B |
| 🟩 NVIDIA A100 40 GB | llama.cpp · CUDA | ✅ Qwen3.6-27B-MTP @ 128 k · gemma-4-31B · 🟡 Qwen-AgentWorld-35B, GLM-5.2 |
| 🖥️ Linux/macOS CPU | llama.cpp | ✅ Qwen2.5-0.5B (CI smoke) |

Ran a combo we don't list (NVIDIA, AMD/ROCm, …)? A one-line PR note is gold. ❤️

## Troubleshooting

<details><summary>NVIDIA GPU idle / running on CPU</summary>

`llama-cpp-python` was built CPU-only. Re-run the **installer** (not bare pip),
then check: `python -c "import llama_cpp,glob,pathlib as p; print(glob.glob(str(p.Path(llama_cpp.__file__).parent/'lib'/'*cuda*')) or 'NO CUDA')"`
</details>

<details><summary>Context feels small / <code>llama_decode returned 1</code></summary>

Fixed — context is auto-sized to the model and GPU (no flat cap), and a full KV
ends the turn cleanly so you can continue. Force it with `KAS_CTX` / `KAS_CTX_MAX`.
</details>

<details><summary><code>Failed to create llama_context</code> on load</summary>

The KV didn't fit VRAM; kas auto-backs-off, or cap it: `KAS_CTX_MAX=32768`. Keep
Flash Attention on for sliding-window models (gemma).
</details>

<details><summary>Garbage output / leaked markers (<code>&lt;|im_end|&gt;</code>, <code>&lt;turn|&gt;</code>)</summary>

GGUF-backend bugs, fixed — upgrade `kas`. A new model leaking a marker is a
dialect gap; open an issue with the model id.
</details>

## Other backends

Point the agent at any Anthropic/OpenAI-compatible server (Ollama, LM Studio,
vLLM, TGI): `kas --base-url http://host:port`. For the full feature set (KV
continuation, per-session caches), add an adapter under `server/backends/`
satisfying the `EngineLike` port in
[`server/core/ports.py`](server/core/ports.py). A native Ollama/LM Studio adapter
is a great first PR.

kas

kas 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

kas

kas 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README