Hi All - I’ve been exploring what’s possible with agentic tools running locally on the DGX Spark and am sharing my latest experiment: Mica, a workspace where you interactively collaborate with multiple agents — voice, coding, and general — to build and run apps.
Compared to other coding agents, Mica helps you and agents curate and manage context for a project on a shared Canvas, instead of primarily within a chat session. This is useful when building apps, when you and the coding agent(s) need to align on scope, functionality, dependencies, and implementation approach.
A few things I’m exploring with Mica:
-
Project materials stay as plain files that any tool can read or modify, while Mica renders them as rich Cards that make the meaning and structure easier to work with.
-
Specs, scope, dependencies, implementation notes, and outputs stay visible and adjust as the project evolves.
-
Runnable Cards as the output: dashboards, visualizers, document tools, workflow UIs, and scripts.
-
Multiple local agents working alongside local voice input/output. Current setup: Qwen3.6-35B, Parakeet STT, Kokoro TTS, OpenCode, Qwen Code, and Claude Code agents.
-
Modern local inference in the loop: vLLM, NVFP4, speculative decoding, and llama-server support.
-
A local-first workflow that can run on DGX-class hardware, smaller GPU setups, or cloud endpoints when needed.
-
Long-running project workspaces that stay available across devices, so you can keep multiple agent builds moving in parallel and drop back in from a laptop, tablet, or phone.
Mica runs in Docker and comes set up for both vLLM and llama-server inference. You can also point it to cloud endpoints when you want to compare models or use cloud inference.
Check it out here: https://github.com/mica-agent/mica
Cheers,
Rob