A coding setup that remembers

Most AI coding setups reset every session. Mine does not. It is a four host Claude Code platform where the agents share one identity, one written set of operating rules, and a memory that survives machines and model upgrades.

• Orchestration. MCP integrations for mail, calendar, cloud, and local services; a model router that sends routine work to a 12 GB local GPU first and only escalates to cloud models when the task earns it.
• Memory. Persistent agent memory in TypeScript and Bun on SQLite full text search plus a vector index. Every session is extracted into structured facts; scheduled recovery catches anything the live extraction missed.
• Code quality. Lifecycle hooks enforce the rules the agents cannot be trusted to remember: secret scanning before commits, pre push guards that refuse the wrong remote, read before edit, check state before act, fail loudly.
• Why it matters to a customer. It is the reason a discovery call becomes a demo, a sizing, and a deck in about a day and a half instead of a week.

Agents need a system of record

Every business that deploys agents is going to be asked the same question an auditor asks about any other system: who did what, when, and why. Most agent deployments cannot answer it. This one can.

• Every session is committed to git with its full transcript, the decisions made, and the approaches rejected, then indexed for keyword and semantic search. Any answer an agent gave can be traced back months later, with context.
• Least privilege by design. Agents get the minimum authority to complete a task; outbound actions (email, messages, pushes) require an explicit human word, and the platform refuses them otherwise.
• Evaluation before trust. An evaluation workbench compares outputs against explicit rubrics, flags critical failures and coverage gaps, and separates real behavior drift from noise.
• The thesis. The audit trail is not a compliance afterthought. It is what makes an agent's output usable as evidence in a business decision.

Build spec in, live site out

A reusable, AI assisted delivery workflow that turns a structured build spec into a fully live production website with no manual steps: repository, deployment, DNS, and content, end to end. It has shipped more than ten production sites, including a 104 page site in a single multi agent run with zero errors, and more than 30 desktop plugins.

• Repeatable. The same pipeline that ships a technical field guide ships an interactive customer demo, so the second one costs a fraction of the first.
• Verified, not assumed. Every build is screenshot checked at real sizes before it is committed; nothing ships unread.
• In the day job. Customer demos and prototypes in about a day and a half, built with the same tooling.

Fourteen instruments, no server

Practical tools for designing and evaluating AI applications. Everything runs in the browser; what you type stays on your device, and there is no server to send it to, which is the point when the inputs are sensitive.

• Prompt laboratory, token and cost planner, context packer, retrieval lab, model selector.
• Evaluation workbench, workflow decomposer, agent designer, permission planner.
• Latency budgeter, failure investigator, drift monitor, signal tester, AI stack mapper.

Linux desktop, custom development

More than 30 published desktop plugins for the Omarchy Linux desktop and over 120 public repositories: hardware control, media, messaging bridges, and system monitoring, written in the open with contributors and users.