Inspiration
The difficult part of a project handoff is often the change between updates: a different owner, a moved deadline, a new blocker, or a commitment that quietly disappears from the list. A fresh summary alone can hide that history.
Commitment Brief gives a reviewer a short path from “what changed?” to a follow-up they have actually read. It is designed for small project teams that already maintain structured commitment records and need a clearer review before the next meeting.
What it does
The browser companion compares two supplied commitment snapshots. It separates additions, removals and changes, and shows before-and-after owner, status, due date and title. The MCP comparison retains full before-and-after evidence; the browser evidence pane inspects the selected snapshot’s supplied quotes. A removed item is explicitly treated as absent from the newer snapshot, not as completed.
The reviewer can see which commitments need attention on a chosen review date, inspect exact quotes supplied with a commitment, and create a follow-up preview. They can edit the wording and approve that version to enable a local Markdown export link. The browser handles file saving; completed delivery must be checked on the user’s device. Approval is a review step in the browser; the application has no message-sending tool.
The included example is fictional. A launch budget moves from Maya to Theo, its due date changes, and its status becomes blocked. The reviewer can inspect both snapshots before deciding what to ask Theo.
How we built it
The core is a deterministic MCP service with four tools: compare_snapshots, summarize_commitments, inspect_evidence and preview_follow_up. The browser companion calls the tools over Streamable HTTP. A separate Node.js server and SDK client provide a reproducible local implementation and test route.
We used the official MCP TypeScript SDK's client, server and Node packages, with Zod schemas for normalized snapshots. Actual SDK-client integration tests verify the 2025-11-25 initialization and list/call workflow, as well as the SDK's newer protocol negotiation. These are real protocol calls, not prewritten results presented as tool activity.
The server compares supplied fields; it does not extract commitments from arbitrary notes. Status, ownership, dates and quotes come from the caller. Quotes are preserved exactly, but are not independently verified against an external source. Later review dates compare supplied due dates and do not establish whether work has since finished.
This entry uses the Alexa+ track's working MCP-server route. It has no live Alexa connection, voice interaction, AI-model invocation or AWS-service integration. We are not claiming the AWS Builder mini challenge.
Challenges we ran into
Changes must remain understandable without inventing facts. The comparison matches stable commitment IDs, returns full before-and-after fields and does not equate a missing item with completion. Empty snapshots are valid, while an unknown requested owner or commitment is an explicit error. The comparison rejects a later snapshot dated earlier than its baseline.
Protocol compatibility needed direct evidence. The SDK supports a newer negotiation path, so we exercised the required 2025-11-25 path separately with an actual client. For the standalone local server, we added fixed loopback binding, exact Host/Origin checks and a bounded request body. These are local server checks, not a production-security certification.
The Worker build also required a narrow browser-export alias for pkce-challenge. The hosted browser endpoint was moved to /api/mcp after /mcp failed to reach the application on the hosting platform; all four hosted actions then worked. Detailed reproduction and feedback are included in the repository.
Accomplishments
Fifteen standalone integration tests and two Worker-adapter tests pass. The tests cover four-tool discovery and calls, attention states, snapshot changes, exact evidence, empty inputs, caller-declared data labels, freshness, chronology, unsent previews and local HTTP safeguards. Type checking, application lint, production build and dependency checks pass.
The product connects the review steps: inspect a change, see the supplied evidence, decide what needs attention, then control the wording of the next follow-up. Editing the draft clears its approval. The export payload is verified; automatic downloads in our in-app test browser were canceled, so completed user-device download remains a manual check and is not claimed in the video.
What we learned
A reliable review tool needs more than a convenient summary. The snapshot date, the meaning of a removal and the source of every status claim affect what a person should do next. Those boundaries need to be visible in the interface and tool results.
What's next
The next useful extension would be an explicitly authorized adapter from a team's existing commitment tracker, with clear import provenance and refresh dates. We would test the workflow with project leads before making claims about time saved or customer adoption. External messaging or live Alexa integration would be separate additions with their own implementation and validation.
Prior work and assistance
Commitment Brief builds on a problem explored in our separate Commitment Desk project. This new implementation was created during this hackathon window and adds a separate four-tool MCP service, before-and-after snapshot comparison, and a new browser review workflow. It does not reuse the earlier project's Bedrock demonstration as evidence for this entry. The earlier submission remains separate.
Development, documentation and testing used AI coding assistance. All example project data and the video's narration are synthetic. The replacement video is a continuous capture of the functioning browser companion on a Mac, with no intentional timing cuts; the encoded output is resampled to 30 fps. It demonstrates actual MCP tool calls, draft editing and approval reset. It does not demonstrate a live Alexa connection or a completed file download.
Judge access
The public demo needs no login or API key. The repository includes installation instructions, the standalone SDK demonstration and tests. No Alexa or AWS account, paid model key, or external project-tracker connection is required.
Log in or sign up for Devpost to join the conversation.