Lighting: drive real Windows browser/apps as text refs (~97% fewer tokens)
At a glance
Lighting (MIT, Windows) exposes your real logged-in browser and desktop apps to Claude Code / Codex plugins or any MCP client as compact text element refs instead of screenshots — claimed ~440 tokens for a page look vs ~15,850 with screenshots (~97% less), one-line click results (~13–53 tokens). Self-learning routines; 135/135 selftest badge on the repo. Alternative to vision-heavy computer-use loops when you already have local logins. This workflow is listed under computer-use, Windows, token-efficiency, MCP, browser-automation.
SuperCompress v2 compresses context against the query (keeps answer-critical lines verbatim): ~64% mean cut on their B5 coding-agent suite with 24/24 evidence-retention passes; MCP/plugin for Cursor/Claude Code/Codex plus hosted API and open weights.
HN (Ted Haley): Pensieve locally indexes code, docs, and past agent sessions (watches repos/worktrees/uncommitted edits), exposes MCP search so Claude Code/Codex/Cursor get the few relevant passages with file:line instead of multi-round grep; Qwen3-Embedding-0.6B + SQLite/trigram + turbovec; macOS-focused OSS.
Show HN: `npx subagent-reuse --setup` adds an MCP that scans ~/.claude/projects/, builds SHA-256 Merkle trees of files each subagent already read/modified, and route_task decides REUSE vs CREATE_NEW before spawn—structural overlap only, no embeddings; ~1k LOC JS + 52 tests.
Rust MCP server makes Claude Code the team lead over Claude/Codex/Gemini workers; Stop-hook + asyncRewake pushes replies (~50ms) into the lead session, with a local live team UI.