A technical manual for terminal-fluent operators. Build a local-first agent system with memory, tools, and real containment — on hardware you control.
Run the stack on hardware you control, with explicit boundaries and no mandatory third-party runtime.
Go beyond prompts: build the loop, memory model, tool contract, autonomy layer, and containment perimeter.
Every chapter is written to be run, checked, and verified from first principles.
Install the inference stack, configure the system prompt, and validate the end-to-end path with a raw API call.
Build the core loop that separates an agent from a one-shot response generator.
Split behavioral directives, runtime context, and operator context into stable configuration files.
Add conversation state, working memory, and vector-based long-term recall.
Introduce Janitor, Vault, Ghost-Write, and the permission gate that controls tool execution.
Run the system on systemd timers with standing orders, exit semantics, and notification paths.
Add process isolation, filesystem restrictions, network boundaries, and audit logging.
Deploy the whole system in stages and operate it with explicit trust boundaries.
Context assembly, inference, response parsing, observation, reflection.
Sliding context, working digests, and vector retrieval.
JSON-over-STDIO tools with explicit permissioning.
Autonomy through process isolation and network boundaries.
Sovereignty Access Pack
One-time payment. Zero recurring fees.