Inspiration
Planning dinner rarely happens in one sitting. You invite six people. Later, someone mentions an allergy. By Friday, you want to know what is ready and what still needs doing, without explaining the whole evening again.
That is the small, everyday problem behind HestiaRelay. A helpful assistant should carry a plan forward as life adds details, while leaving the decisions with the people who will live with them.
What it does
The demo starts with a simple request: “We're having six people over Friday evening, budget $120.” HestiaRelay saves the plan and creates a preparation list.
In a new session: “Remember Ana is allergic to nuts.” The detail joins the existing plan. The guest count and budget do not need to be entered again.
In a third session: “Are we ready for Friday?” HestiaRelay brings back the plan, the people, the constraint and the unfinished work. Its timeline shows how those pieces arrived in different conversations. Restarting the server keeps them.
Decisions stay with the plan too. A sensitive proposal shows its exact details before approval or rejection. If the household context changes, an earlier approval cannot authorize the new proposal. In this prototype, approval records consent only: no purchase, payment, message or account change is carried out. Allergy information is a planning constraint, not a food-safety guarantee.
How it works
The guided browser experience and a real Streamable HTTP MCP server use the same Python service and SQLite database. A new session sends only its new message; the service retrieves the saved context. The video finishes with a real MCP client reading the same household shown in the browser.
The planner has two explicit sources. The demonstration uses deterministic
planning, so anyone can reproduce it without an AWS account. The optional
Amazon Bedrock adapter calls converse through boto3. One Nova Micro call was
tested in temporary AWS CodeBuild on September 12, 2026, with its response and
usage recorded. That historical cloud test is separate from the video.
This is an Alexa+ experience simulation, with working MCP integration. It is not connected to live Alexa+. AgentCore and Strands are not integrated.
The hardest part
Remembering a fact is only part of continuity. The harder question is what that fact changes. A new dietary constraint can make an old proposal unsuitable; an approval must not quietly carry over. HestiaRelay stores the proposal's context and checks it when a decision is made.
The same care applies after a restart. The demo closes the actual server process and starts another against the same database. The household and its decisions return together. Tests also cover rejection, changed context, failed planner calls and backup restoration.
What is working
The public MIT project includes the simulator, MCP server, a non-root Docker setup and instructions for trying the three-session story. Validation covers 156 tests with 98.14% application coverage, real MCP clients, desktop and mobile Chromium, container recreation and authenticated recovery checks. Detailed results and security limits are linked below.
The demo shows how you can return to the same plan without surrendering control over what happens next.
What comes next
The first audience is people who coordinate meals and small household events. The next product test is straightforward: can someone return to a plan, spot the unfinished work and understand what they approved without repeating the earlier context? Time savings and fewer mistakes are hypotheses to test, not results already measured.
The prototype currently supports guided English messages and one fictional household. Broader conversation, screen-reader testing and household isolation are the next steps before a public multi-user service. Any future shopping or messaging integration would need its own consent-bound execution adapter.
Submission references
- Primary track: Alexa+. Mini-challenges: Open Source and AWS Builder.
- Watch the public narrated demonstration.
- Public MIT source and run instructions.
- GitHub username: Elinfiny. Contribution: implementation PR #3, release PR #15.
- Open Source contribution: built a shared household service for the browser and MCP, persistent context and an exact-consent ledger, with tests and run instructions so other developers can reproduce and extend the pattern.
- Built with: Python, MCP Python SDK, SQLite, Pydantic, Starlette, Uvicorn, HTML/CSS/JavaScript, Amazon Bedrock, boto3, Docker and GitHub Actions. The full development-tool feedback and temporary AWS services are listed in product feedback.
- Development used AI assistance through ChatGPT/Codex. The repository preserves the implementation, review history and reproducible checks; generated advice is never treated as permission to act.
- Judge evidence map, friction log, release audit, historical Bedrock proof.
Log in or sign up for Devpost to join the conversation.