ToDoVex

The suggestion action sends a project's existing tasks and its name to GPT-3.5 Turbo and asks for five missing items as structured JSON. TodoVex filters duplicates, creates the tasks, and tags them with a system AI label so they're easy to spot. It's a cheap model call, and the output usually reads like a decent colleague's checklist rather than a finished plan.

Search converts your query into an embedding and runs a vector search over your own tasks, returning up to 16 close matches. Results follow meaning, so synonyms and fuzzy descriptions still hit. The filter scopes every query to your user ID, which keeps one account's tasks out of another's results.

Each task carries a name, optional description, due date, priority and label, and can hold subtasks one level deep. Priorities run five steps from lowest to urgent, and completed tasks move to their own list instead of vanishing. Anyone coming from Todoist will recognize the shape on the first screen.

Projects group work, labels cut across them, and Today plus Upcoming collect whatever is due soon. Convex pushes changes to every open tab in real time, so a checkbox you tick on a laptop updates a phone browser without a refresh. System projects and system labels sit alongside your own and can't be repurposed.

Authentication runs through Auth.js v5 and a Google OAuth flow, with sessions stored in the Convex database. Every query and mutation checks the signed in user before touching a row, which is the part most tutorial apps skip. The tradeoff is that a Google account is the only way in.

The repository ships the full application: schema, queries, actions, components and configuration. Point it at your own Convex deployment, add an OpenAI key for the AI features, and you have a private task manager. Forking is explicitly allowed, and several forks already strip the AI layer to run fully offline.