Researched and written with 309,550 tokens (in 10 · out 30,778 · cache-read 65,545 · cache-create 213,217) across 5 Claude calls

Coding agents

• anomalyco/opencode: open-source terminal coding agent, installable via curl, npm, or scoop; the main OSS alternative to Claude Code and Codex, and many tools in this batch target it.
• anthropics/claude-code: the reference terminal agent; the README now marks npm install as deprecated in favour of native installers, so update install scripts and docs.
• anakitaonrails/ai-memory: shared long-term memory that lets you quit Claude Code mid-task and resume in Codex in the same directory without re-explaining; it targets vendor lock-in at the context layer.
• No pull requests, no code review. We push to main.: Thorsten Ball describes agents that read production logs, watch deploys, grant feature flags, and verify their own work end to end; the guardrails are the interesting part.
• Sphinx: Git hook that quizzes you on your agent's code before it ships, using minimal tokens via your existing agent or any OpenAI-compatible endpoint; HN skeptics say the people who'd benefit will click through anyway.

LLM mechanics

• Gemini is 50% more expensive than it should be: Theo reports a 27% cache miss rate on Gemini's prompt cache, which would erase most of the caching discount; check your own cache-hit telemetry before trusting list prices.

Orchestration and multi-agent

• stablyai/orca: runs Codex, Claude Code, OpenCode, or Pi side by side, each in its own git worktree, with mobile notifications and follow-ups; worktree-per-agent is becoming the standard isolation pattern.
• mattpocock/sandcastle: TypeScript sandcastle.run() runs agents in Docker, Podman, or Vercel sandboxes with a branch strategy and merges commits back; this is programmable sandboxed orchestration.
• Orchestrator-V2 in T3 Code: a user reports Claude driving Grok CLI and Devin-hosted Gemini as subagents; cross-vendor subagents are showing up in mainstream tooling.

Agentic SaaS and tooling

• NitroStack: build, test, and deploy MCP servers in one place; it's relevant if you're shipping MCP tooling.
• Ultracite v8: linter and formatter pitched at agentic development, meaning fast, opinionated, and auto-fixing for agent output.
• CrewTower: talk to Claude Code from the MacBook notch; a niche UI novelty.
• Landing page to live product demo skill: a Claude Code skill where a scripted cursor and camera drive your real UI instead of a screen recording, so the demo stays in sync with shipped features.

Vibe-coding and repos

• SEO tools for coding agents list: a curated, auto-refreshed list of 247 open-source SEO and GEO (AI-citation) repos usable from Claude Code, Codex, Cursor, and OpenCode, filtered from 651 candidates.
• Claude Code black hole renderer: Claude Code built a physically accurate raw-WebGPU renderer and checked its own physics, a good example of an agent verifying its output against a domain spec.