😸 Caps Lock got a promotion

gm legends, happy Saturday.

Today: a Caps Lock key with a new job, a voice for your coding agents, and a teleprompter that actually listens to what you're saying instead of just scrolling at you. Plus a Microsoft approval saga with a twist ending.

Capsomnia turns your Mac's Caps Lock key into a physical switch that keeps the machine awake with the lid closed, so an AI agent or SSH session doesn't die the moment you close it. Taketo Fujimaki (@fuji_mak) built it free and open source, no account, no telemetry.

🔥 Our Take: This already has 2.3 million views and 6,000 downloads in Japan, and it hasn't properly launched anywhere else yet. Amphetamine already does the keep-your-Mac-awake thing, this is the only one that turns a physical key into the switch, with the Caps Lock light doubling as the status indicator. A tool this small usually needs a pitch, but 6,000 downloads before it even launched here says it doesn't need one.

Heard listens to your Claude Code, Codex, and Cursor sessions and turns the important parts into speech, one voice per agent, summarized at the project level instead of five terminals talking over each other. Kelly Sun (@itskellysun) built it after two years on the other half of talking to agents: they answer back in scrolling text you have to sit and stare at.

🔥 Our Take: Sled already turns Claude Code into a voice, Heard's actual point is holding context across a whole session so it knows when a boring-looking stretch is about to need a decision, not just reading you a wall of tool-call output. Permission prompts and failures always interrupt no matter what, everything else gets judged against what you've already heard. The real test isn't the voice, it's whether it stays quiet at the right moments.

Inference should feel simple. Once AI hits production traffic, you're really managing routing, retries, observability, GPU bills running while nobody's home.

DigitalOcean Serverless Inference handles all of that in one API — OpenAI- and Anthropic-compatible, 55+ models including Claude, DeepSeek, Qwen, and Llama. Scale to zero when you're not using it. Pay per token when you are. VPC and zero data retention on by default.

230 tokens/sec on DeepSeek V3.2 — 3.9× faster than AWS Bedrock (Artificial Analysis). Runs on the same bill as your databases and Kubernetes. One less thing to reconcile.

Speechius listens to what you're actually saying and scrolls your script to match, on-device, one-time purchase, no subscription. Jay Imdahl (@syntacops) has presented on Zoom and YouTube almost daily since 2020 and spent two years annoyed that every teleprompter forced a choice: follow a fixed scroll speed, or lose your script entirely.

🔥 Our Take: Every other teleprompter makes you pick a lane: fixed scroll speed you chase, or no script at all once you start improvising. Jay's presented daily since 2020, so he lived that gap for years before building the thing that listens to your actual words instead of your volume and adjusts. FlowPrompter and TalkEze already do speaker notes, neither one's actually listening to you.