A personal workbench for running AI coding/control agents on my own hardware, with host-side credential isolation, egress policy, and vendored MCP bridges. Each sub-directory is an independent experiment or deployment — they share the repo for convenience, not because they compose into a single product.
nemoclaw_agent/— NemoClaw (Claude Code in a sandbox) setup: native WSL2 workflow, Discord gateway proxy, bruiser policies, onboarding scripts. Seenemoclaw_agent/AGENTS.md.heb/— MCP server bridging Claude Code to H-E-B grocery ordering. Host-side OAuth, vendored SDK overlay, cart + checkout tools. Seeheb/sources.ironcurtain_agent/— Experimental agent harness around the provos/ironcurtain Signal bridge (submodule).resonantos/— ResonantOS fork integration (submodule →dkaygithub/resonantos-alpha) wired up with an openclaw launcher.heb-sdk-unofficial/— Submodule: fork ofiHildy/heb-sdk-unofficial, tracked on thedev/stackbranch (octopus merge of all in-flight PR branches). See AGENTS.md for the stacked-PR workflow.design/— Architecture notes and three candidate setups (ResonantOS / LangGraph / NemoClaw) covering supply chain, input security, cognitive alignment, authorization, infra, observability, and human-in-the-loop.dbg/— Ad-hoc debugging scripts, policy YAMLs, patches, and one-off tooling for the above. Nothing here is production — it's the scratch drawer.
openclaw/ and pi-mono/ are local working copies of upstream projects kept outside this repo's history. .har captures are excluded so request traces with live cookies/tokens never get committed.
git clone --recurse-submodules https://github.com/dkaygithub/local_agent.git
Active, evolving, and deliberately messy. Not a product. Use at your own risk — no license is attached yet, so no usage rights are granted.