Hearsay

Hearsay turns what you say into work. Bee (on an Apple Watch or a Bee device) transcribes what you say during the day. Hearsay listens to that stream, extracts the commitments buried in it — "after lunch I need to fix the login redirect bug", "remind me to email Professor Kim tomorrow at ten" — and turns each one into a Bee todo, a GitHub issue, or a note, with a macOS notification so you know it landed. A companion Claude Code skill lets your coding agent pick those issues up with the spoken context attached.

Built for the Bee track of the Build, Ship, Shape: Amazon Developer Hackathon (October 2026).

How it works

Apple Watch (Bee app) ──► Bee cloud transcription
                                 │
                                 ▼  bee stream --json --types new-utterance,...
                          hearsay watch
                                 │  buffers utterances per conversation
                                 │  (20 s of silence, or 6 utterances)
                                 ▼
                      LLM intent extraction  (claude -p │ Gemini │ fake)
                                 │  {type: issue|todo|note, title, due, quote}
                 ┌───────────────┼──────────────────┐
                 ▼               ▼                  ▼
        bee todos create   gh issue create     ~/.hearsay/log.md
                 └───────── macOS notification ─────┘

Everything the project knows about you comes from Bee at runtime: the live stream (bee stream), recent conversations (bee now), and the todo API (bee todos). Nothing is mocked in normal use.

Requirements

• Node.js 20+ (tested on 24), no npm dependencies
• Bee CLI: npm i -g @beeai/cli, then bee login (enable Developer Mode in the Bee iOS app first: Settings → tap Version five times)
• One LLM backend:
  • claude (default): the Claude Code CLI on your PATH, logged in
  • gemini: GEMINI_API_KEY set
  • fake: keyword heuristics, offline, for tests and demos
• Optional: GitHub CLI logged in, for issues

Install

git clone https://github.com/1jsjs/hearsay && cd hearsay
npm link            # puts `hearsay` on your PATH
npm test

Usage

# Live: listen to Bee and act on what you say
hearsay watch --repo 1jsjs/my-project

# Same, but only print what it would do
hearsay watch --dry-run

# Catch up on the last two hours of conversations (no live stream needed)
hearsay replay --minutes 120

# Try the extractor on a sentence
hearsay extract "After lunch I need to fix the login redirect bug"

# Run the full pipeline against a bundled sample stream, offline
HEARSAY_LLM=fake hearsay watch --dry-run --input fixtures/live-shape.jsonl --window 1

Option / env	Meaning	Default
--repo, HEARSAY_REPO	GitHub repo (owner/name) that receives issues. Without it, issues become Bee todos	—
--window	seconds of silence before a conversation window is processed	20
--max	utterances that force processing early	6
--input	JSONL file to replay instead of the live stream	—
HEARSAY_LLM	claude, gemini or fake	claude
HEARSAY_HOME	where state, raw events and the log live	~/.hearsay

State (state.json) remembers which utterances were handled and the last 50 titles, so restarting or replaying never files the same task twice. Every raw stream event is appended to events.jsonl for debugging.

Claude Code skill

skills/hearsay/SKILL.md teaches Claude Code to read your recent Bee context through the Bee MCP server (bee mcp connect claude-code) and start on the issues Hearsay filed. See that file for setup.

Privacy

Hearsay only reads your own Bee account through the official CLI. Transcripts are sent to the LLM backend you choose for extraction and are otherwise kept locally under ~/.hearsay. Tell the people around you that Bee is listening.

License

MIT