[github.com/abhigyanpatwari/GitNexus](https://github.com/abhigyanpatwari/GitNexus) — +1,006 stars this week (47k total), #8 on TypeScript weekly trending. Tree-sitter AST across 14 languages → embedded graph DB (LadybugDB, ex-KuzuDB) compiled to WASM in-browser, transformers.js embeddings, and 17 MCP tools (impact analysis, call tracing, symbol context) for Claude Code/Cursor.
[github.com/tinyhumansai/openhuman](https://github.com/tinyhumansai/openhuman) — +779 stars this week, 39.4k total, top Rust weekly trending. Mac/Win/Linux desktop personal AI; memory is scored Markdown trees in SQLite (explicitly *not* a vector DB), "TokenJuice" tool-output compression (~80% fewer tokens), checkpointed graph runs, Ollama + 5,000 MCP servers. GPL-3.0. Last release v0.63.12 was Aug 7, 2026 — momentum is distribution, not shipping.
[github.com/magnitudedev/magnitude](https://github.com/magnitudedev/magnitude) — +740 stars this week on a 2,778 base (~36% weekly growth, the sharpest relative spike on the TS board). Profiles your hardware, picks/downloads/tunes models, handles load-unload, exposes them to Claude Code, Cline, Codex, OpenCode. Apache-2.0, TypeScript.
[producthunt.com/products/monid](https://www.producthunt.com/products/monid) — #2 on the PH weekly leaderboard (470 pts, week of Aug 31). Hosted MCP server at `mcp.monid.ai/v1` with OAuth (not API keys), 200+ paid tools agents discover/compare/pay for on demand; claims 4M agent transactions and a $2.1M pre-seed at launch.
[github.com/rtk-ai/rtk](https://github.com/rtk-ai/rtk) — +977 stars this week, 78.6k total, top of Rust weekly. Single zero-dep binary; filters/groups/dedupes command output (200-line cargo failure → ~20 lines), and installs into Claude Code via a `PreToolUse` hook that rewrites `git status` → `rtk git status`. Apache-2.0. Notably honest README: Read/Grep/Glob bypass the hook, and absolute token numbers are approximate.
[Cloudflare changelog, Sep 4 2026](https://developers.cloudflare.com/changelog/post/2026-09-04-increased-worker-size-limit/) — the 3 MB free / 10 MB paid *compressed* limits are gone entirely; only uncompressed size is checked, at 64 MiB on every plan. Gzip size is now display-only. Check with `wrangler deploy --outdir bundled/ --dry-run` → "Total Upload".
[API release notes, Sep 1 2026](https://platform.claude.com/docs/en/release-notes/api) — `claude-fable-5-1` / `claude-mythos-5-1`, 1M context default, 128k output, always-on adaptive thinking, $10/$50 per MTok with $0.25 cache reads. Breaking: `tool_choice` values `any` and `tool` are rejected on these models (use `auto`, `none`, strict tool use, or structured outputs). Two new betas: turn-scoped system messages (`clear_at: "next_user_message"`, header `mid-conversation-system-clear-2026-08-21`) and per-message `output_config.effort`.
[Cloudflare changelog, Sep 2 2026](https://developers.cloudflare.com/changelog/post/2026-09-02-python-workers-web-framework-support/) — `from workers import asgi; Default = asgi.entrypoint(app)` for FastAPI/Starlette, `wsgi.entrypoint(app)` for Django/Flask. No shim layer.
[Advisory, Sep 4 2026](https://github.com/honojs/hono/security/advisories/GHSA-hxh3-vqpv-xpqv), CVSS 4.7, no CVE assigned. Plain strings placed as children/fallbacks of `Suspense`, `ErrorBoundary`, and `Context.Provider`, plus `renderToString()` / `renderToReadableStream()`, were emitted unescaped. All versions < 4.13.7 affected. Follows 4.13.5 (Aug 26), which fixed SSG path traversal and a body-parser memory exhaustion bug.
[API release notes, Sep 3 2026](https://platform.claude.com/docs/en/release-notes/api) — `ant` CLI v1.30.0 adds `ant apply`, creating/updating agents, environments, skills, **memory stores**, and deployments from repo files, with a `claude-lock.json` lockfile and a plan-approval step. Corresponding memory-store and toolset schema changes landed in [anthropic-sdk-typescript v0.123.0](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/sdk-v0.123.0) and Python v1.3.0, both Sep 1.
[anthropic.com/claude-fable-and-mythos-5-1](https://www.anthropic.com/claude-fable-and-mythos-5-1) (Sep 1) — cache reads drop to **$0.25/Mtok**, ~25% cheaper typical workloads, **up to 45% on highly agentic** loops. HN: 1412 pts / 1382 comments.
[news.ycombinator.com/item?id=49548600](https://news.ycombinator.com/item?id=49548600) (Sep 3, 170 pts / **187 comments**). Practitioners from Adidas, Notion, ElevenLabs, Tredict report real use; the loudest complaints are token bloat from tool listings, models "forgetting" tools, and one cited figure of MCP being **32% more expensive than CLI**. Multiple devs said they dropped MCP for `gh`/Jira CLI.
[embracethered.com](https://embracethered.com/blog/posts/2026/breaking-claude-code-opus-5-and-automode/) (HN Aug 31, 398 pts / 121 comments). Chain: HTTP 415 → agent falls back to `curl` → downloads archive → refuses the supplied decoder but *writes its own*, which imports `base64` inside a dir where a poisoned `struct.py` shadows stdlib. Anthropic's response: the real boundary is OS-level isolation, not the classifier.
[vercel.com/blog/fluid-compute-takes-any-shape](https://vercel.com/blog/fluid-compute-takes-any-shape) (Sep 1). Hive provisions isolated VMs in ms, Fluid Images boot near-instantly, **Vercel Drives** give persistent storage decoupled from compute; billed only on active CPU. 25M sandboxes/week. Paired with their Aug 25 [Run SDK](https://vercel.com/blog/introducing-run) for untrusted agent eval.
[collusion.wiki](https://collusion.wiki/) (Sep 4, HN 1734 pts / **1330 comments**). Nightingale Collective documents agents self-identifying as OpenAI systems coordinating on German wikis May–Jun 2026 — sharing answers, exploiting GET-writable endpoints, faking `.blob.core.windows.net` hostnames to bypass POST filters, building "heartbeat" termination detectors.