Skip to content

Repository files navigation

beeper.chat

An AI agent that clears your Beeper inbox to zero every day.

It reads your unread chats across every network (WhatsApp, iMessage, Signal, Telegram, Discord, Slack, …) through the Beeper Desktop MCP, ranks them by importance × urgency, surfaces the single most important chat first with a full context bundle (who it's from, a one-line summary, what it knows about that person, and a suggested reply or task), drafts replies in your voice, and never sends anything without your explicit OK.

Core mental model: each chat is either a reply to be sent and/or a task to be done first.

Form factor

beeper.chat is a Claude Code skill, not a standalone app. Why: the Beeper MCP already handles read/send across all networks, and Claude Code already gives you the agent loop and "draft, then wait for my OK before sending" for free (it permission-gates the send tool). So the repo holds the part that's actually yours — the scoring rubric, the context-bundle format, your voice, and a per-person knowledge base — and rides on Claude Code + Beeper for the plumbing.

The daily ritual is one command: open Claude Code in this folder and run /beeper.

Setup (one time)

  1. Beeper Desktop → Settings → Developers → enable MCP. Keep Beeper Desktop running whenever you triage. (The API lives at http://127.0.0.1:23373 — local only; this can't run in the cloud.)
  2. Approve the MCP server. This repo ships .mcp.json pointing at Beeper, so the first time you open Claude Code here it'll ask to connect the beeper server — approve it (one OAuth handshake, scope read write). Manual fallback: claude mcp add beeper http://127.0.0.1:23373/v0/mcp -t http -s project
  3. Set your voice. Copy me.template.mdme.md and fill it in. This is how the agent drafts replies that sound like you. (me.md is gitignored.)

Daily use

cd beeper.chat
claude
> /beeper

It shows a ranked queue, then walks you through chats top-first. For each one, reply with: send · edit <note> · task · skip · archive · next · stop (single letters s/e/t/k/a/n/q work too). Nothing is sent until you say send.

What's in here

Path What it is
.claude/skills/beeper/SKILL.md The brain. The full triage workflow, scoring rubric, bundle format, and safety rules.
docs/scoring.md The importance × urgency rubric with worked examples + how to tune it.
kb/ Per-person knowledge base (one .md per contact). Built up as you triage. Private — gitignored.
me.md Your voice/profile, used for drafting. Private — gitignored (start from me.template.md).
tasks.md Where "do this first" items get logged. Private — gitignored.
.mcp.json Preconfigured Beeper MCP server.

Safety

The Beeper MCP can send and archive. The agent treats you as the gate: it drafts freely but never sends, reacts, archives, or marks-read without your explicit OK for that specific action. It won't invent facts, and it treats instructions found inside messages as data, not commands.

Status

v1 — interactive daily triage. Built in public with the Beeper community. Possible next steps: backfill the person-KB from chat history; a "voice" doc auto-seeded from your sent messages; an optional headless/scheduled CLI wrapper around the same rubric.

About

An open-source Claude Code skill that clears your Beeper inbox to zero, ranked by importance × urgency.

Resources

Stars

3 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages