Sidekick

A voice-first AI partner that watches movies with you, plays games, and handles code, git and computer tasks through one open MCP server, asking approval before anything risky.

Comment

Inspiration

Voice assistants, browsers and coding agents each live in their own silo. I wanted one partner that could sit with me on the couch during a movie, play a quick game, and also help with real work like running tests or committing code, without giving up control over anything risky. That became Sidekick.

What it does

Sidekick is a voice-first AI partner with one open MCP server behind it.

As a companion

  • Movie Buddy: say "Hey Sidekick, what just happened?" while a film plays. It lowers the video volume and answers using only the real subtitle lines, so it doesn't invent plot details.
  • Voice games: chess (full rules via chess.js with an AI opponent), Tic-Tac-Toe and 2048, all playable by voice.

As a helper Sidekick exposes 11 tools through MCP: coding.execute, coding.run_tests, workspace.inspect, vcs.diff, vcs.commit, system.info, system.volume, system.media_control, system.open, system.clipboard and web.fetch.

Safety by design Every tool has a risk level. Reading and running tests happen automatically. Editing files and committing wait for the user's approval.

One server, many front ends The same tools work from an Alexa+ style voice flow (Alexa skill endpoint plus an Alexa+ simulator in the web app), from the React web app, and from external MCP clients like Claude Desktop, Cursor or MCP Inspector.

How we built it

  • Core brain: Python and FastAPI for intent routing and Alexa request handling
  • Execution service: Python tool registry with an approval flow and live job events over SSE
  • MCP server: JSON-RPC 2.0 over stdio [plus Streamable HTTP, spec 2025-11-25 - keep only if finished and tested]
  • Coding providers: Amazon Bedrock (Qwen3 Coder), an HTTP API provider (Claude/OpenAI-compatible) and a local CLI provider
  • Web client: React 18, Vite and TypeScript, with chess.js for the chess engine
  • Movie Buddy: an OpenAI-compatible endpoint (Groq), grounded on subtitle lines
  • Hosting: Render, https://sidekick-jq3l.onrender.com/
  • Quality: 395 backend tests (pytest) and 179 frontend tests (vitest)

Challenges we ran into

  • Deciding which actions run automatically and which need approval, and making that clear to the user
  • Keeping Movie Buddy honest by grounding every reply in real subtitle lines instead of model guesses
  • Sharing one tool registry across stdio, HTTP and the Alexa path without duplicating logic

Accomplishments that we're proud of

  • One tool registry that serves Alexa, the web app and external MCP clients
  • An approval flow tied to each tool's risk level
  • A companion side (Movie Buddy, voice games) and a work side (coding, git, system) in the same product
  • 395 backend and 179 frontend tests, and an MIT-licensed open-source repo

What we learned

how mcp works and alexa routing decision and making user experience smooth and good

What's next for Sidekick

  • Publish a full Alexa+ experience once the platform allows it
  • More tools (calendar, notes) under the same risk and approval model
  • Per-user permission profiles and an audit log of approved actions
  • More Movie Buddy features, such as watching with friends in sync

Built With

Share this project:

Updates

Submission history