LLM Tasks is a task and session layer built specifically for agent-driven software work. It keeps durable projects, tasks, subtasks, dependencies, labels, sequence ordering, task events, session logs, handoff notes, progress history, and next-task suggestions in PostgreSQL so agents have a stable work plan across conversations.
The system is designed around the difference between transient in-chat checklists and durable project work. Agents can start an explicit tracked session, record progress against existing tasks, close the session with blockers, decisions, warnings, next steps, and handoff notes, then resume later from the last known project state instead of reconstructing the plan from memory.
Recent work added a usable board around the MCP layer: a Kanban-style web UI on port 8766, live updates through SSE, archived task views, parent navigation for subtasks, project editing, cross-project task moves with label-drop reporting, manual sequence reordering, project-scoped labels, and task events that include session log entries.
The MCP surface exposes prompt-oriented workflows such as `/blueprint`, `/tasks`, `/next`, `/progress`, `/session`, `/handoff`, and `/triage`. Behind those prompts are ranking and sequencing rules that account for priority, staleness, complexity, labels, and dependency unlocks, with `task_triage_apply` intentionally separated so an agent can show recommendations before changing the project plan.
Operationally, it shares the llm-memory Postgres container, installs Codex and Claude integrations, can launch an embedded web board with the MCP server, and writes a user systemd service so the local task board can come back after a reboot. The result is a developer tool for preserving momentum on multi-session builds rather than just listing todos.
