Three GitHub repos crossed 2,500 stars this month and — by accident — they slot together into a deployable AI services stack. Here is how I am running them on one terminal.
The Three-Layer Stack
| Layer | Repo | Stars | What It Does |
|---|---|---|---|
| Visual | neilsonnn/image-blaster | 2,944 | Photo → explorable 3D world in ~5 minutes (Claude Code × World Labs marble-1.1 × FAL hunyuan-3d × ElevenLabs SFX) |
| Front-end | nexu-io/html-anything | 2,599 | Agentic HTML editor — auto-detects 8 coding CLIs, 75 skills × 9 surfaces, zero API keys required |
| Memory | Tencent/TencentDB-Agent-Memory | 2,530 | 4-tier local memory (L0–L3), v0.3.4 released May 13 2026 |
Why the Memory Tier Matters
Tencent's published benchmarks for the L0–L3 hierarchy versus flat RAG:
| Metric | Flat RAG → 4-Tier |
|---|---|
| WideSearch | 33% → 50% |
| Tokens used | −61.38% |
| PersonaMem | 48% → 76% |
That is roughly half the tokens for double the persona recall — which is the difference between a demo and a paid product.
Install in One Terminal Session
git clone https://github.com/neilsonnn/image-blaster && cd image-blaster && bun install
git clone https://github.com/nexu-io/html-anything && cd html-anything && pnpm install
git clone https://github.com/Tencent/TencentDB-Agent-Memory && cd TencentDB-Agent-Memory && uv sync
No glue code required — html-anything auto-detects whichever CLI is on your PATH, and TencentDB-Agent-Memory exposes an HTTP shim that image-blaster can call as a memory backend.
Monetization Angles
-
Visual-first agency — sell branded 3D walk-throughs (real estate, product launches) using
image-blaster+html-anythinglanding pages. - Long-running personal agents — license a private assistant with the 4-tier memory that actually remembers clients between sessions.
-
White-label AI dashboards —
html-anything's 75 skills × 9 surfaces map onto every common SaaS dashboard pattern.
Walk-through (≈60 seconds):
More stack breakdowns: gptaiclips.com · prev episode: gptaiclips.com/v/ep57
