In public preview since Oct 28, Max runs several independent review agents on models from different providers and has an aggregation agent reconcile them into one review; GitHub's docs estimate about $10 in AI credits per Max review versus $1.50 for Balanced. Copilot approvals are now GA and count toward required approvals, but get dismissed when new commits land (docs).
Shown on HN Nov 1 under Apache 2.0, npm install g sweet search , then sweet search init and sweet search index , gives your agent six fully local search, grep, trace and read tools plus a tuned prompt, with the index updating itself as you edit. On the author's own 200 task SWE rebench run it cut agent cost 5–33% with the same number of tasks solved, but it's macOS and Linux only (Windows via WSL2) and costs GPT models about 2.5 points of retrieval accuracy…
Shown on HN Oct 31, adding @khalilgharbaoui/opencode claude code plugin to your OpenCode 1.x or 2.x config drives your logged in claude CLI as a subprocess, so you get Claude's own tools, MCP servers and skills (which OpenCode 2.2.3's built in Claude Agent provider skips) while OpenCode still runs Bash, Edit and Write behind its own permission prompts. It never touches your OAuth token and bills however your CLI does, but it's a brand new one person project (Show HN)…
Shared on HN Oct 31, the zero dependency agent reads and edits files, runs shell and git with checkpoints and undo, speaks MCP, loads SKILL.md skills and works with any model in the models.dev catalog, using about 1.7 MB of memory on Windows. Only Windows is fully tested (Linux builds and runs, macOS is source only), and its source available license bans commercial use and public forks…
Posted Oct 30, each runtime session gets an authenticated URL that forwards HTTP, SSE and WebSocket traffic to 1–5 allowlisted ports (8000–12000) and stops working when the session ends. It's microVM runtimes only, with public HTTPS and no PrivateLink for inbound connections…
Released Nov 1 under Nace.AI Open RAIL M, Drex Max returns a probability for every option in one forward pass instead of generating text, so routing, triage and guardrail checks can run locally: a 4 bit build fits 24 GB Apple Silicon Macs, with a Q8 0 GGUF and llama.cpp and Ollama forks for 48 GB machines or CPU. Its 67.52 on Decision Index 0.4 is self reported, and local Mac and CPU paths are text only for now…
Spotted something we missed? Start a thread.
