Scalably’s Sept 19 field report: two RTX 5090s ran Unsloth’s Qwen3.8-27B NVFP4 via vLLM 0.27.0 for 14 days of real agent traffic—82.6% prefix-cache hits, 28k requests, zero engine errors. A faster MoE lost the job on tool recovery (1/20 vs 12/20), so they kept the dense model.
Posted to HN Nov 4, Maki uses a tree sitter index tool to skim file skeletons before reads, a sandboxed code execution tool that filters tool output before it reaches context, and subagents that pick a weak, medium or strong model per task, plus a bash permission check that parses git diff && rm rf / as two separate commands. It runs on Anthropic and OpenAI API keys, xAI, Gemini, Copilot, Ollama, llama.cpp, OpenRouter and more, and its headless JSON output matches Claude Code's…
Shown on HN Nov 4 under Apache 2.0, npm i g ravenry app starts a board on localhost:4600 where you can answer questions, pick menu options or attach files for sessions running in cmux or iTerm2, with per project briefs and an automatic hand off once a session passes half its context. It adds two hooks to your Claude Code (and optionally Codex) settings after asking, runs only while its terminal is open, and phone access works on your home Wi Fi only (Show HN)…
Shown on HN Nov 4, pip install AutoCua gives one Python and Rust agent that switches between computer use on macOS, Windows or Linux, a real terminal, your own Chrome through an extension, and iOS or Android devices, picking numbered elements from the accessibility tree plus OCR instead of guessing pixel coordinates. It supports 48 models across 9 providers and sends no telemetry beyond the model provider you choose (Show HN)…
Shown on HN Nov 4, npx yes run spaceapp@latest install sets up a Docker stack on Linux, macOS or Windows 11 where OpenCode, Codex, Claude Code, Gemini, Qwen, Kimi, Grok and DeepSeek CLIs run on your own accounts with shared skills, MCP servers and pgvector memory, plus an optional coordinating agent that only asks you about releases, deletes, money and direction. Agents see only directories you register, and the docs say rooms are not a security boundary between untrusted users…
Shown on HN Nov 4, npm install g @one ai/one code gives you subagents, Git worktrees, plan and auto modes, permissions and context management on Anthropic, OpenAI, Gemini, OpenRouter or local models, reading your existing CLAUDE.md , skills and .claude/ config with no import step. You can switch models mid session or give each subagent its own provider, and features tied to Claude Code's hosted or desktop services are out of scope (Show HN)…
Spotted something we missed? Start a thread.