A fully local AI agent stack for a support-engineer workflow on Apple Silicon (48 GB): Hermes CLI + llama.cpp + Qwen3.6-35B-A3B + in-place control of the real managed Chrome.
Work tools enforce device posture, so the browser is driven where it already
is — no automation profile, no debug port. See docs/runbook.md Phase 2.
No cloud inference. No data leaves the machine. This repo is the public
bootstrap artifact — clone it on the target machine and follow the spec.
Machine-specific configuration (site adapters, selectors, runbooks) lives in
a git-ignored local/ directory and is never committed here.
- Design spec: docs/superpowers/specs/2026-07-10-work-agent-design.md
- Work-machine runbook: docs/runbook.md — the ordered setup + verification procedure
| Component | Role |
|---|---|
llama.cpp (llama-server) |
Local inference server, OpenAI-compatible on :8080. Replaced LM Studio — every runtime choice is an explicit flag (config/llama-server.md) |
| Qwen3.6-35B-A3B-MTP GGUF (Q4_K_M) + mmproj | The model — MoE, fast per-step latency in long agent loops. GGUF, not MLX (see setup/02-model.sh for why). MTP head gives self-speculative decoding |
| Hermes CLI | Agent harness — skills, memory, subagents; both coding and operating |
cua-driver via hermes computer_use |
The only browser path. Drives the real managed Chrome in place — accessibility tree, DOM via Apple Events, pixels last. Background: no cursor movement, no focus steal |