Yesterday's Top Launches: 5 Tools from October 29, 2026

Five new developer tools and apps landed yesterday, and the mix is unusually wide: a Chrome extension for cleaning up YouTube captions, an open-source agent training platform, a calm mobile app for intentions, a self-hosted machine-per-agent platform, and a local medical image viewer. Some are more mature than others, but each one solves a specific irritation rather than chasing a trend.

subtitle.im

Burned-in subtitles on YouTube are one of those small annoyances that never quite gets fixed. subtitle.im takes a different angle than most caption tools: it manages the captions that already exist, letting you choose whether captions follow YouTube’s setting, stay on, or stay off across videos. Local OCR detects burned-in subtitles so it can handle duplicate soft captions according to your preferences.

The scope is deliberately narrow. It doesn’t generate, translate, download, or remove burned-in subtitles, which will disappoint anyone hoping for a full caption workflow. What it does offer is a free trial of ten successfully handled videos without signup, with an optional Pro tier adding channel rules, video exceptions, and memory of confirmed closures. It’s a TypeScript-built desktop tool for people who watch a lot of YouTube and are tired of toggling captions manually.

Overmind

The pitch here is that you don’t need a frontier model for every task. Overmind is a training platform for specialized agents, and it walks teams through a full loop: connect your repository and production traces, inspect where an agent failed, build datasets from real traces in a notebook-style Data Workshop, evaluate candidate prompts and models against actual production tasks, then fine-tune an open model on your own data.

The enterprise angle is serious. It’s open source under AGPL-3.0, self-hostable with Docker Compose, and offers zero third-party training, meaning your operational data never trains public foundation models. That matters for teams whose traces contain sensitive business detail. The founders’ backgrounds in defense intelligence, ML infrastructure, and fintech scaling are cited as the foundation.

Where I’d push back: the platform assumes you already have production agent traffic worth learning from. Early-stage teams with a handful of traces won’t get much from the data workshop or evaluation gates yet. But for teams running agents at real volume and wondering why they fail in production, the decision-level inspection view alone looks useful. Access starts through a free Console account.

Savor

Most task managers assume you already know what you’re doing. Savor bets the opposite: many things people carry around aren’t tasks yet, they’re partial thoughts and possibilities. The app calls itself an Intention Layer, which is marketing speak, but the underlying idea holds up. You capture by voice, text, or photo, and it shapes the fragment into a card with what it is, when it matters, and what to do next.

Cards live in three lanes (Now, Soon, Later), reminders are gentle and configurable, and there’s a beta Take Action feature that helps draft plans and words. Card handoffs keep delegation explicit: the other person takes ownership, gets the reminders, and you see a read-only copy until you take it back. Privacy gets real attention, with no training on your captures and age-aware guardrails for teen accounts.

The deliberate absence of streaks, badges, and dashboards is the whole point. That will read as relief to some and as missing structure to others. It’s free on iPhone and Android while monetization is paused, so trying it costs nothing.

Paddock

This one solves a problem that’s about to get very common: giving coding agents real work without handing them your laptop. Paddock creates a box for each agent, a Linux machine with its own Ubuntu XFCE desktop, files, tools, and Docker daemon, all running on a VPS you own. Installation is a single script that sets up the server, HTTPS, and Docker.

The details are thoughtful. Network policy can be allow-all, deny-all, or an allowlist of domains and CIDRs, changeable on a running box. Cloud metadata is always blocked. Snapshots save every minute and on stop, so an agent that breaks its environment gets rolled back instead of rebuilt. Drives under /mnt outlive boxes and can mount read-write to one box or read-only to many. You drive it from a CLI, dashboard, HTTP API, or TypeScript and Python SDKs.

Claude Code, Codex, OpenCode, Pi, Prime, and Kimi come preinstalled and can be sent tasks with paddock prompt. It’s MIT licensed and free, though you pay for the VPS. The tradeoff is obvious: you’re now administering a server. For teams already running agents in earnest, that’s a reasonable swap.

OpenRadView

A free, open-source desktop viewer for your own medical images. It opens DICOM, NIfTI-1/2, and ZIP archives locally, with no account, no telemetry, and no scan uploads. That last point is the real draw, since scan CDs and ZIP files are awkward to inspect without sending them somewhere.

The feature set goes well beyond basic viewing: window and level controls, linked MPR with axial, coronal, and sagittal planes sharing one crosshair, oblique planes, experimental 3D and maximum intensity projection views, and distance and angle measurements where geometry is known. Three demo scans ship with the app so you can try every view before opening your own files. There’s also an optional MCP bridge so an agent you already use can see a series you explicitly allow, revocable at any time.

The disclaimers are everywhere, and rightly so. It’s for education and personal research, not diagnosis, and it displays images without interpreting them. The 3D views are experimental with a 512 MiB budget and GPU limits. If you have your own scans and want a local way to explore them, this is a solid option, but pair it with a clinician for anything that matters.

Quick links

Yesterday’s batch leans toward infrastructure and privacy rather than flashy interfaces. Paddock and Overmind both assume agents are already doing real work and need somewhere serious to run or improve, while subtitle.im, Savor, and OpenRadView each carve out a small, specific job and do it locally. Worth a look depending on which problem you’re living with.