Glossary

Node — a visual element on the org-chart canvas representing a role. Each node is backed by an autonomous AI agent.

Role — the identity a node carries (CEO, CTO, CMO, Personal Assistant, etc.). Shapes the agent’s domain and decision biases.

Workspace — a top-level container for an org. Holds nodes, skills, artifacts, channels, and activity history. Equivalent to a “company” or “project”.

Skill — a markdown workbook with YAML frontmatter + step-by-step instructions the agent follows to execute a task. Stored at data/knowledge/<workspace_id>/skills/<slug>.md.

Agent — the runtime backing a node. A process (typically a systemd-nspawn container) that runs the Claude SDK, looks up skills, follows instructions, produces artifacts.

Coordinator — the in-container service (http://10.101.0.1:9000) agents reach back through to POST milestones, fetch instructions, and message other agents.

Workbook — the XLSX file a skill uses to persist structured data (sourced leads, qualified leads, daily metrics, etc.). Pre-created by the system with fixed sheet schemas.

Starting context — a permanent system-prompt addition for a node, set once when you create the node or via settings. Affects every future response.

Activity timeline — the chronological event log for a node (org-activity). Every thought, tool call, decision, milestone, artifact lands here.

Artifact — a file (spreadsheet, PDF, draft, screenshot) a skill produces. Stored per-workspace in the artifacts shelf.

Decision — a commitment an agent makes during a skill, logged to the workspace’s decisions log. Can require request_approval to advance.

Milestone — a significant checkpoint in a skill (e.g. “ICP defined”, “campaigns launched”). Visible on the node’s card.

Channel — an inbound + outbound communication path for a node (WhatsApp, Discord, Slack, email).

MCP — the Model Context Protocol , an open standard for AI tools. Direktor’s MCP server lets Claude Desktop invoke node skills.

MCP token — a per-node bearer (of_dkmcp_*) that authorizes a Claude Desktop client to drive a specific node.

Best-effort fallback — when no skill matches a request, the agent uses general reasoning. If it succeeds, the result is persisted as a new learned_by: agent skill.

skill_learned event — the stream event that fires when the agent completes a task without a pre-existing skill, persisting what it learned as a new skill markdown.