Omni Apps

A personal project I keep coming back to. Most agent tooling asks you to write instructions in a markdown file and hope the model follows them, which falls apart as soon as you change the model.

Omni Apps has the agent write the work as code instead. Validated workflow graphs are stored in PostgreSQL and compiled into LangGraph at run time, and users review and edit them through streamed A2UI components. A workflow becomes repeatable and re-runnable regardless of which model fills each step.

It has a dual memory store in PostgreSQL that pairs pgvector semantic recall over high-dimensional embeddings with an entity-relation knowledge graph the agent enriches after every turn. The platform learns a user's recurring subjects and how they connect, so it carries context between conversations and needs less instruction on the tenth workflow than on the first.

A modular MCP connector layer surfaces external tools, such as mail, calendar, documents, a live browser and databases, to the graph-builder agent, so a tool is connected once and reused across every workflow. It follows the connector principles of the Linux Foundation's agentgateway project, runs on Google Cloud, and deploys entirely through infrastructure as code.