Remventory is a single Go application for a self-hosted, agent-first inventory prototype. The app should center Remy, use Postgres for durable state, expose MCP tools as the stable capability boundary, and keep all data-changing agent actions behind proposals that the user approves or rejects.
- Go for the server, persistence, MCP, agent integration, and web backend.
- Postgres for users, categories, attributes, items, proposals, settings, and agent events.
github.com/mark3labs/mcp-gofor the MCP server.- Google ADK for Go for Remy. The current Go module import path is
google.golang.org/adk, even though the upstream repository isgithub.com/google/adk-go. - OpenAI-compatible model configuration for local or remote LLMs.
- Trial
github.com/ag-ui-protocol/ag-uifor structured agent/UI events. If the repository does not expose a Go package, keep local event payloads protocol-shaped and document the gap. - One Docker image for the application container.
- Do not commit category, item, or duplicate-resolution changes directly from agent output. Create a proposal first; commit only through an explicit confirmation path.
- Keep the web UI Remy-centered. Do not make a spreadsheet clone or a serial chat transcript the primary output.
- Keep category definitions structured and item-specific values in JSONB.
- Do not introduce RAG, embeddings, vector search, file uploads, full auth, or multi-user collaboration in the prototype.
- Documentation should explain setup, configuration, and usage. Avoid documenting internal details that are obvious from code.
- Should categories eventually support subcategories?
- What access-control model should replace the optional prototype token?
- How long should agent events and proposal history be retained?
- How should future image/file assets be exposed once S3-compatible storage is added?