Covalent.ai

Inspiration

We were tired of writing massive prompts for every single thing we do. Every AI agent today waits for you to ask—but that's not how humans work. We juggle contexts, repeat workflows, and constantly switch tasks. We asked ourselves: What if AI could learn your patterns and act before you even ask? That's when Covalent.ai was born—a desktop agent that watches, learns, and proactively suggests actions, turning AI from a tool you use into a partner that works alongside you.

What it does

Covalent.ai is a promptless AI desktop agent that continuously monitors your screen, learns your workflows, and proactively suggests actions.

Here's how it works:

Silent Learning: Covalent runs in the background, capturing screen content and building a semantic understanding of what you're doing Smart Context Engine: Our proprietary graph-based system organizes everything into a hierarchical knowledge tree, using embeddings to instantly retrieve relevant context from your entire work history Proactive Intelligence: Instead of waiting for prompts, Covalent anticipates your next move—drafting emails, scheduling meetings, or pulling up documents before you ask Dual Execution: Actions execute via MCP tool calls (for GSuite, Slack, etc.) or autonomous screen control for anything else Built on Tauri + React with a Python Flask backend, powered by Claude Sonnet 4.5 for action generation and Gemini embeddings for context retrieval.

How we built it

Frontend: Tauri + React + TypeScript for a lightweight, native desktop experience with real-time screen monitoring

Context Engine:

SQLite database storing a hierarchical graph of nodes (tasks/projects) Google Gemini embeddings (text-embedding-004) for semantic vectorization Cosine similarity-based retrieval to find the most relevant context Recursive parent-child relationships to maintain context hierarchy Action Orchestration:

Flask API endpoints for screen ingestion and action triggering Claude Sonnet 4.5 for proactive action generation based on current context LangGraph workflow to route tasks between MCPs and screen controller Composio integration for GSuite operations Screen Control: Custom automation layer for executing actions that fall outside API boundaries

Challenges we ran into

Being able to figure out exactly what a user is doing at all times and learning from that is not trivial. Coming up with algorithms and workflows to automate these efficiently was the challenge