GitHub Weekly Vol.4: 6 Projects Worth a Look This Week

★ 44,896·CSS·read 11-02

A design skill for Claude Code, Codex, Cursor and OpenCode: fifty material themes, each with a finished-page picture, layouts built around what the page must do, and browser-rendered gates that block machine-made tells. MIT.

Why look

Topping weekly trending; a theme is a material, not a layout template — two prompts in one theme yield two different pages; gates come in tiers, and the hard floor (contrast, focus, reduced motion, invented facts, phone layout) never ships.

The catch

Browser-rendered checks (--render) and screenshots need puppeteer-core and Chrome installed separately; this run covered static checks only, and the fifty themes were not each verified.

Best for

Anyone letting agents build pages who is tired of the same AI look; for a single detail, start with audit.

Open on GitHub →

★ 99,352·Rust·read 11-02

A design vocabulary for coding agents: 24 commands (bolder, polish, critique, audit and more) steer changes, plus a detector built on 62 deterministic rules that scans with no LLM and no API key. Apache-2.0.

Why look

The detector is this issue's most concrete piece: on the same sloppy page it flagged 1.5:1 low contrast, overused Inter and the purple AI palette in one pass; a browser live mode swaps variants in place and writes back to source.

The catch

Some of the 62 rules depend on a DESIGN.md design-system file — projects without one only get the generic rules; live mode edits local source, so review the validation script it runs first.

Best for

Anyone iterating on an existing page (critique first, then change); anyone building from scratch should read the docs' warning that new-page mode is the harder one.

Open on GitHub →

★ 109,316·TypeScript·read 11-02

Persistent memory across agent sessions: lifecycle hooks capture tool-use observations, compress them into summaries stored in SQLite, and MCP search tools retrieve them in three layers (index, timeline, details) across Claude Code, Codex, OpenCode and more. Apache-2.0.

Why look

The right idea — memory follows the project, not the tool — with one store searchable from several agent tools; the three-layer retrieval is a genuinely token-saving design: index first, details only on demand.

The catch

Not tested this issue: default install signs in via browser for a hosted-memory trial, and full session capture needs a real agent session; a local provider skips sign-in, but the end-to-end path was not run.

Best for

Anyone switching agent tools inside one long-lived project; anyone wary of memory leaving the machine should check the local provider and private tags first.

Open on GitHub →

★ 12,793·Week +2,410·Rust·read 11-02

An open-source parametric 3D CAD app in pure Rust: constrained sketches, a feature timeline (extrude, fillet, shell and more), assemblies and sheet metal, with desktop, browser and headless CLI running the same commands, plus an MCP server for agent modeling. MIT or Apache-2.0.

Why look

69 releases in under a month since creation, and the author reports 75 of 76 public reference parts matching analytic or FreeCAD-computed volume, area and centre of mass (author's figures, not independently recomputed); the no-install browser build is a real on-ramp.

The catch

Not tested this issue: the desktop app needs a graphical environment, the source build is a large Rust workspace, and this box has no Cargo toolchain, so the CLI modeling path was not run; Windows installers are currently unsigned.

Best for

Anyone looking for an open CAD alternative, or wanting agents to output STEP/STL directly, in hardware and robotics; verify precision beyond the 76 reference parts before production use.

Open on GitHub →

★ 13,571·Rust·read 11-02

A local-first, cross-platform open CapCut replacement: multi-track cutting, auto-captions, TTS and voice cloning, background removal and 100 effects, all offline, plus a built-in MCP server and JSON-RPC API so agents and scripts can cut video too. AGPL-3.0.

Why look

“Editing as script” is the real differentiator: one project file across macOS, Windows, Linux and Android, with AI models downloaded once from Settings and offline after; this issue covered code and docs only.

The catch

Not tested this issue: still beta, source builds need Rust 1.93+ and FFmpeg 7+ dev libraries, this box has no Cargo, and no end-to-end desktop edit was run; AGPL-3.0 is unfriendly to commercial integration — read the licence before commercial use.

Best for

Short-video and tutorial makers who refuse subscriptions and uploads; anyone wiring editing into automation should start with the MCP and JSON-RPC docs.

Open on GitHub →

★ 2,433·Rust·read 11-02

An archive extractor that works from the tail and truncates the original as it goes: ZIP, 7z, tar, RAR and split archives, reusing space freed by the archive for the extracted files, with a desktop app and CLI. MIT or Apache-2.0.

Why look

It solves a real problem: big archives that cannot unpack on a nearly full disk; in this run 3 files came out byte-identical and the archive was removed; it does the space math before touching anything and refuses when space is short.

The catch

The default 2 GB free-space reserve refuses first on small disks until --keep-free is lowered; non-interactive runs need --yes, piped y does not count as confirmation; Windows and macOS installers are currently unsigned.

Best for

Anyone whose disk is always nearly full and who unpacks big game or dataset archives; ordinary unpacking just needs --keep-archive.

Open on GitHub →