Atlas

Atlas is my self-hosted personal operating agent built on OpenClaw. Instead of treating an AI system as a generic chatbot, the project defines it as a bounded operator with an explicit identity, clear tool permissions, a durable memory model, and a narrow set of approved surfaces for work.

At the center is a modular governance pack under Governance/, with files such as AGENTS.md, BOOTSTRAP.md, HEARTBEAT.md, IDENTITY.md, SOUL.md, TOOLS.md, and USER.md controlling startup, policy enforcement, operating cadence, and change management. Local memory is split between a compact startup profile in Governance/MEMORY.md and dated append-only memory files, while shared writing, data, research, analysis, and document artifacts live under stable project workspaces. Runtime-configured web search handles live retrieval, AgentMail enables tightly constrained email handling through a trusted allowlist, and optional trusted WhatsApp, image, transcription, and voice surfaces are governed explicitly rather than assumed.

The system is designed to stay natural at the interface while being much more disciplined underneath. Atlas can move between conversation, recall, research, synthesis, drafting, analysis, artifact creation, and delegated execution, but it does so inside an explicit governance model rather than through vague prompt instructions or unchecked autonomy. Internally, the system uses one public capability, atlas_companion, with private helper roles for evidence gathering (albert), writing (oscar), bounded execution (carlos), premium deep research (marco), and difficult fallback work (francis), so it can delegate without turning into a noisy multi-bot interface.

One example of that operating model is the local wiki workflow, which compiles immutable raw source files into an Obsidian-compatible knowledge base with source pages, concept pages, entity pages, maps, indexes, and maintenance logs. Other local skills cover notes, posts, editorial work, digests, and site-specific publishing flows. The system stays local-first and portable: wiki pages use project-relative source links rather than machine-specific absolute paths, so the resulting knowledge base can move between machines.

The goal is practical leverage without black-box behavior: better recall, better synthesis, and more useful delegation, while keeping the system local-first, auditable, and intentionally constrained.