Global cache · 4-level cascade

Firmware builds,
cached at the edge.

A shared cache server for ESP32, STM32, and RP2040 projects. Cuts generation time with a 4-level cascade from byte-match to Claude AI.

Admin Login Projects
L0–L3
Cache Levels
RS256
JWT Auth
MLX
Local LoRA
mTLS
Secure by default

Everything your firmware team needs

One server handles caching, discovery, and generation — no cloud required.

🔌

mDNS Discovery

Servers announce themselves on the LAN. No config files, no hostnames — they just appear.

🔒

TLS Certificates

Each node generates a keypair and CSR on first boot and requests a cert from the MakeGPT CA.

⚡

MinIO Object Store

L0 cache backed by local S3-compatible storage. Byte-exact hits return in milliseconds.

🤖

Claude AI fallback

L3 calls the Claude API when no local cache or model can answer — fully logged and billed to your key.

4-level cache cascade

Each request waterfalls through levels, stopping at the first hit.

L0

Byte Cache

Exact-match lookup in MinIO. Fastest possible hit.

L1

Vector Search

Semantic similarity via Qdrant — handles near-duplicate prompts.

L2

Local LoRA

Fine-tuned model via MLX on Apple Silicon. No internet needed.

L3

Claude API

Full generation via Anthropic. Result is written back to L0–L2.