My interactive graph, a snapshot from Sep 21, 2026.
Releases
Jul 2026v1.3.0: Memory Canvas and Input graph.
Jun 2026v1.2.0: On-device Qwen3-ASR transcription, system-audio capture, trainable voice ID for speakers, and automatic sorting of events into project folders.
May 2026v1.0.0: First release, with local capture of screen, audio, and typing.
Sparky-toolbox
A personal agent harness built for the NVIDIA DGX Spark.
It is optimized specifically for local models. The agent lives in an offline sandbox, every tool call passes through a gateway with budgets, hooks and approvals, and untrusted content is read by isolated subagents, so the model stays precise and reliable across long, multi-step tasks.
One agent run traced through its loops, beside a workflow task.
It allocates the machine to the work. Each task, subagent and tool call is given the model, memory and context suited to it: models are loaded and swapped per job, tools run in containers within their declared limits, and context is assembled on demand instead of all at once.
How Sparky-toolbox extends itself.
It also improves itself. When the agent lacks an ability, it designs a new tool, task or subagent, builds and tests it in an isolated dev area, and ships it once approved. Its capabilities are not limited to what it was given.
