I posted an earlier version of Personal Task Assistant as an open-source queue for human/AI work. The project has moved from "show agent-ready tasks" to a more complete handoff loop: - FastAPI UI/API - SQLite local setup, Firestore support for hosted deployment - atomic claim endpoint for Codex-owned tasks - idempotent ingest with external_id - MCP server with 9 tools for Claude/Codex-style clients - deterministic eval process for the MCP tool surface - AGENTS.md + skills so future agents follow the project contract - daily digest after explicit schedule installation - optional Claude work loop The daily worker is deliberately constrained: it only receives the task_assistant_* MCP tools, has no shell/repo/web, cannot mark tasks done, and can only send completed work back to waiting_review. The product principle is: humans decide, agents execute, review stays explicit. Development continues beyond 0.6.0. I will share the next changes separately instead of mixing several releases into one post. I would appreciate feedback on the safety boundary and the task model. Repo: https://github.com/J3d1-fm/Perso...

Personal Task Assistant

Bounded daily AI worker for your task queue

Bounded daily AI worker for your task queue