glimmer-local-agent

glimmer-local-agent is a full work agent that runs entirely on your own Mac. People who would gain most from an AI assistant often can't use one: confidential documents, client email and medical forms can't be pasted into a cloud chatbot. This agent keeps every model call on the machine and shows you what leaves it. You ask in plain language. The agent answers, works through mail, calendar and files, or drives a real Chrome window. On browser steps, Muse Glimmer 30B reads the set-of-marks screenshot locally and writes a short grounding note: the elements that matter and where they are, what changed, and whether the goal looks done. A local Qwen planner reads the page text plus that note (never the pixels) and decides. Each model runs on its own local Ollama instance. Anything consequential (send, submit, pay, delete) stops at an approval card, every connector call is audit-logged, and a Private-mode panel shows live, with a network-egress meter, that no model process talks to the internet. We measured instead of claiming. On 45 real-web tasks run twice (90 runs per arm), with a fresh conversation per task, the same frozen harness and three scorers: - Qwen planning + Glimmer grounding every screenshot: 82/90 verified (78 frozen, 80 full-answer). Right on 82 of 84 finished runs; confidently wrong on 2 runs instead of 8. - Qwen alone: 80/90 (76 frozen, 78 full-answer). The cost is speed: median 3.5 vs 1.9 min per run (684 vs 363 task-minutes). Everything ran on one Apple M5 Max with 128 GB of memory: no cloud, no API calls. The browser engine underneath is the author's private prior work (built before the hackathon); everything built for the hackathon is open source under Apache-2.0.