Drawgent: drive Claude Code and Codex from a live Excalidraw canvas
A quiet night for models, a loud one for tooling: a canvas that steers your agent, cross-agent memory over MCP, and how DeepSeek runs its sandboxes.

Copy markdown
Drawgent: your agent works on the whiteboard with you
A single Rust binary serves a local Excalidraw editor and bridges to your existing Claude Code, Codex, or opencode session over ACP. Drop an AGENT: note on any shape and it redraws live; it screenshots the canvas to reason about layout. On the HN front page ~7h ago (90 points). Run `drawgent setup claude` then `drawgent up`.
Hemory: an MCP memory layer your coding agents can query
Launched on Product Hunt (318 votes): an always-listening capture app (phone, Apple Watch) that turns conversations into a searchable store and exposes it to Claude Code, Codex, and Cursor over MCP. The pitch is one memory your agents read instead of re-explaining context each session; the always-on mic is the obvious trade-off to weigh.
A Claude Code skill that post-mortems your chess games
Open-source skill (brumar/chess-postmortem-skills, 68 points on HN) that chains vision, Stockfish, and TTS to turn a game into an annotated PGN, an HTML review viewer, and a narrated video. A clean weekend template for wiring a multi-tool Claude Code skill around any analysis pipeline.
DeepSeek details the sandbox behind its agents
A new paper on DeepSeek Elastic Compute describes the platform running roughly 3M sandboxes a day, 380K concurrent, and 5,000+ creations per second. It is research, not a release, but the numbers are a useful reference if you are sizing your own agent-execution infrastructure.