Radar · 2026-09-04

Generated by a daily research loop, unedited. 14 items. Evidence is quoted from the scan; links go to the source.

Trending repos

OpenHuman — local-first personal AI with a Markdown "Memory Tree" instead of a vector DB

[github.com/tinyhumansai/openhuman](https://github.com/tinyhumansai/openhuman) — +939 stars this week (top of GitHub Rust trending weekly), 39.4k total, GPL-3.0. Rust+Tauri desktop; memory is *scored Markdown trees in SQLite mirrored to an Obsidian vault*, explicitly rejecting "vector database black boxes." Rust-enforced local-only mode blocks all cloud calls; Ollama/LM Studio/MLX runtimes, 5,000+ MCP servers.

weave-os/router — sub-50ms per-request model routing in front of Claude Code

[github.com/weave-os/router](https://github.com/weave-os/router) — **+2,082 stars in 7d**, biggest mover on Go trending weekly; 3.8k total, ELv2. Drop-in proxy for Claude Code/Codex/Cursor; routes per *action* not per turn using an in-process ONNX embedder + Avengers-Pro-derived cluster scorer, <50ms, claims 40–70% cost reduction. Anthropic + OpenRouter open models.

GitNexus — MCP-native code knowledge graph, browser + CLI, no server

[github.com/abhigyanpatwari/GitNexus](https://github.com/abhigyanpatwari/GitNexus) — +1,157 stars this week, 47k total. Ships its own embedded graph DB (LadybugDB, WASM in browser), tree-sitter parsing, hybrid BM25 + semantic + reciprocal rank fusion, 16 MCP tools incl. cross-repo analysis, lazy connections evicted after 5 min.

slotstream — 125B MoE on a 48GB Mac by streaming experts from SSD

[github.com/carloslfu/slotstream](https://github.com/carloslfu/slotstream) — Show HN 2026-09-01, 229 pts / 115 comments. Keeps the 3.8GB dense core resident, streams 68GB of routed experts via `pread` into fixed cache slots (not mmap); ~12 tok/s on 48GB MBP. MIT, **Ollama- and OpenAI-API compatible**, Apple Silicon only.

Releases & ecosystem

Claude Fable 5.1 / Mythos 5.1 — cache reads cut 4x

[API changelog, Sept 1 2026](https://platform.claude.com/docs/en/release-notes/api) — 1M context default, 128k max output, always-on adaptive thinking. Cache read drops to **$0.25/MTok (0.025x base, down from 0.1x)** at unchanged $10/$50 base. Breaking: `tool_choice: any`/`tool` now 400s; no way to disable thinking; 30-day retention required.

Anthropic TypeScript SDK v0.123.0 — memory-store + toolset schemas

[sdk-v0.123.0, Sept 1](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/sdk-v0.123.0) — "organization compliance settings, user-profile order_by, memory-store and toolset schema updates." Predecessor [v0.122.0, Aug 27](https://github.com/anthropics/anthropic-sdk-typescript/releases) moved `beta.files`/`beta.skills` to GA shapes and dropped dated beta header pins; v0.121.0 added Standard Schema support for structured outputs and tools.

Ollama v0.33.3 — Gemma4 multimodal on MLX, cached-token reporting

[Ollama releases, Sept 2](https://github.com/ollama/ollama/releases) — Gemma4 gains image and audio support on the MLX engine; **cached prompt tokens now reported**; GGUF default params honored. v0.33.0 (Aug 21) added Claude Desktop gateway support.

Bun v1.4.1 — HTTP/2 in Bun.serve, Postgres decode fixes

[Bun v1.4.1, Sept 4](https://bun.com/blog/bun-v1.4.1) — HTTP/2 via `Bun.serve()` with ALPN; `Bun.write()` streams Response bodies (161 MB → 13 MB peak on a 128 MiB download); `Bun.sql` fixes Postgres `float8`/`float4` text-format decoding and `timestamptz` for pre-1900 dates. 202 issues fixed.

Wrangler 4.129.0 + Durable Objects concurrency raise

[workers-sdk, Sept 3](https://github.com/cloudflare/workers-sdk/releases) — gzip for JSON Pipelines sinks, `--json` for `wrangler workflows`, fix for remote dev with static assets under granular Worker permissions. [Cloudflare changelog](https://developers.cloudflare.com/changelog/): Aug 28, Durable Objects raised from four to ten distinct Dynamic Workers with in-flight requests; Sept 1, D1 free-tier daily query limits now enforced.

Tech-direction conversations

Coding agents pick your dependencies — 17k-run study measures which vendors win

[armature.tech](https://armature.tech/blog/which-tools-coding-agents-install), Sep 3, 226 pts / 104 comments. 5,292 valid sessions, 51 codebases, Claude Code vs Codex vs Cursor. Findings: Stripe wins payments 90%, **Neon wins databases 66%**; mentions ≠ adoption (LangChain cited 194×, chosen 4×). Claude Code searches the web in only ~30% of sessions — it picks from priors. Vendor docs/pricing presentation moved decisions more than technical merit.

Claude Fable 5.1 / Mythos 5.1 — cache reads cut 75% to $0.25/Mtok

Sep 1, 1407 pts / 1375 comments on HN. 1M context, 128K output, base pricing flat at $10/$50, but a cache hit drops from $1.00 → **$0.25 per Mtok — 2.5% of input price**, versus the usual 10% multiplier. Adds mid-conversation effort adjustment (beta). Mythos 5.1 is the same model behind restricted access.

Simultaneous OpenAI + Claude + Grok outage exposes correlated failover

Ask HN, Sep 3, 373 pts / **535 comments**. Root cause traced to a SpaceX Memphis compute center outage hitting rented inference capacity; Cloudflare's CTO disputed the shared-CDN theory. The sharper thread: failover is now universal, so providers fail *in sequence* — tools stampede to the backup without exponential backoff or jitter, and bots amplify far beyond human load.

Local inference on 48GB Macs crosses into agent-usable

Two posts, both Sep 1: ["My local model setup on an M4 Pro Mac Mini"](https://lws.io/blog/my-local-model-setup/) (325 pts / 194 comments) — 48GB, oMLX server, Qwen3.6-35B-A3B 4-bit at 34 tok/s gen / 325 tok/s prompt, with **KV cache persisted to SSD specifically for agent workflows**; and Show HN [slotstream](https://github.com/carloslfu/slotstream) (229 pts), running a 104GB Qwen3.8-Flash-Next on 48GB at ~12 tok/s. Framing: covers "the 80% of requests that don't need Opus."

Cloudflare BotBase for Operators — agents now need declared identity

[blog.cloudflare.com](https://blog.cloudflare.com/botbase-for-operators/), Aug 28. Operators declare purpose, content-use level (Content Signals), and operator type (direct vs. **intermediary**); verification via IP list, reverse DNS, or **Web Bot Auth signatures**. Submissions up ~7× since 2023, forcing automated review.