Backchannel
Backchannel turns WhatsApp into local Markdown and a search index, so you or an agent can find what was said before the follow-up work begins. It builds onwacli.
brew install openclaw/tap/wacli wacli auth # scan the QR once .venv/bin/pip install . backchannel setup
Why this exists
Email archives can be searched, read by other tools and turned back into work. WhatsApp remains where conversations stop.
Decisions, requests and commitments now arrive in group chats. Once they scroll past, you cannot search them, an agent cannot read them, and the follow-up depends on what you remembered. WhatsApp exports one chat at a time, by hand, capped, and without media.
Voice notes make the gap harder to see. You cannot skim one, search one, or give it to an agent to read. In an active group, the decision is often there.
What it produces
--- source: whatsapp type: group name: Studio Reading Group date: 2026-08-07 message_count: 34 has_links: true --- **09:14 — Mara**: did anyone get through the Latour piece **09:22 — Jonas**: half. the middle section is worth it
One file per chat per day. Frontmatter you can filter on, body you can read.
before **14:02 — Priya**: [voice note · 0:47] after **14:02 — Priya**: 🎙 I looked at the venue again and the back room only fits about thirty, so we should decide by Friday either way
Whisper runs on your machine. No audio leaves it, no API key exists.
$ qmd query "what did we decide about the venue" 0.83 Studio Reading Group 2026-08-07 · whatsapp 🎙 …the back room only fits about thirty… 0.71 Ana Ruiz 2026-08-05 · whatsapp …booked it for the 14th
Semantic search over text and transcripts together, offline.
Image OCR is opt-in. The daemon schedules it only when BC_DESCRIBE_ENABLED is set; it is the newest and least-proven capability over long runs.
Source
wacli is a Go CLI that pairs by QR as a linked device, the same slot WhatsApp Web occupies, and mirrors your messages into a local SQLite store. It installs with one Homebrew command and tracks the protocol library weekly. Backchannel reads that store read-only and builds everything else on top.
wacli speaks the protocol through an independent implementation rather than a Meta-built client, which WhatsApp's terms do not permit. Accounts have been blocked for unofficial-client use, and the reported cases cluster around automated sending. Backchannel never sends: there is no send path in the codebase. Low risk, not none, and it is your account.
How it works
The captured store is the only thing that cannot be regenerated. Corpus, transcripts and search index are derivations — delete any of them and run the job again.
Install
Early software. Backchannel installs from a source checkout on macOS. Interfaces, config keys and defaults may still change.
Read this before step one. wacli links as a companion device using an independent implementation of the protocol, not a Meta-built client. WhatsApp's terms don't allow that, so an account could in principle be blocked. Backchannel only ever reads, and the account is yours to risk. Everything below runs on your machine; nothing is sent anywhere.
One Homebrew command. It is the client that talks to WhatsApp; Backchannel reads what it stores.
brew install openclaw/tap/wacliScan the QR from your phone once, under Linked devices. WhatsApp sends roughly three months of recent history straight away; older messages can be requested afterwards.
wacli authFrom a source checkout, create a virtual environment, install, and initialise it on this machine.
python3 -m venv .venv
.venv/bin/pip install .
backchannel setupOne process handles capture, derivation, transcription and indexing on their own schedules. Search needs qmd, which lives on npm.
export BC_WHATSAPP_BACKEND=wacli backchannel daemon npm install -g @tobilu/qmd # for search
Constraints
Linking brings roughly three months of recent history. Going deeper means asking your phone for it, chat by chat, and you can only ever recover what your handset still holds.
Read-only by design. Nothing in it can post a message, which removes a whole class of ways for it to go wrong.
Backchannel reads the archive out. It does not decide anything, create tasks, or reply.
Tested on macOS only. Nothing in the design forbids Linux, but untested is untested and pretending otherwise would waste your afternoon.
Transcription and search run locally. That is a privacy property and also a hardware requirement — Whisper wants a machine that can carry it.