OpenSparX: Skoor 50 (n=3, confidence 0.7, semiconductors)
What they built
MasterAgent is described as a tool to "build AI agents that run 100% on-device," with "sub-100ms latency on Qualcomm NPU" and "zero cloud dependency." [source]
The README describes it as an embeddable, local-first C++17 agent runtime that runs deterministic skills before model inference and validates tool names and arguments. [source]
Version 0.4.0-alpha.3 builds a `sparx` CLI and a C++ SDK, and the README calls it "a small synchronous runtime, not the proprietary durable kernel described by the legacy interfaces." [source]
The demo is an in-memory simulated automotive AC control that needs no model, network or vehicle hardware. [source]
Model inference goes through a separately started llama-server, and optional SQLite storage supports request deduplication and recovery of interrupted tools. [source]
Record
Measured signals (n=3):
Other facts: The repo has 13 forks, and the last push was 2026-09-28. [source] The org has 2 public repos and is not verified, with no website, location or description listed. [source] Founders, team, funding, customers and hiring are not in the evidence.
Inference: 13 forks against 684 stars suggests attention runs ahead of hands-on use.
Why now
Not in the evidence. The only timing data is that the org and repo were created in 2026-08. [source]
Inference: The semiconductor tag rests on the Qualcomm NPU claim. The README describes a CPU-side C++ runtime with release CI on Linux and macOS and does not mention an NPU. [source]
Unknowns
- Who the founders are and what their technical background is: not in the evidence.
- Benchmarks or code supporting "sub-100ms on Qualcomm NPU": not in the evidence.
- What the "proprietary durable kernel" is and who owns it: not in the evidence. [source]
- Commit activity over 90 days: not in the evidence.
- Whether the automotive demo connects to any real customer: not in the evidence.
Inference: The bilingual README (English and Chinese) and automotive demo point to an in-car voice-control use case.
Questions for the room
- The repo claims sub-100ms latency on a Qualcomm NPU, but the README targets Linux and macOS and runs models through llama-server. Show the NPU benchmark: what device, what model, and what is measured?
- The README calls the open code "not the proprietary durable kernel." Who built that kernel, who owns the IP, and is the open runtime meant to sell it?
- The repo gained 684 stars in about two months but has 13 forks. Who outside the team has built on it, and is any automotive company using it?
What would move the Skoor
It would rise with reproducible NPU latency results, named technical founders, or a named design partner. It would fall if the NPU claim has no code behind it, or if pushes stay stalled after 2026-09-28.