v1.3.0-beta.2

Fixes the WhatsApp connection dropping about every 50 minutes, and the
ingestion that stopped a few days later because of it. The tools and the panel
are those of 1.3.0-beta.1.

Fixed

• The connection no longer drops every 50 minutes. The official Evolution
  Go 0.7.2 loses the acknowledgement of a media Status when it reconnects, and
  WhatsApp then ends the connection about every 50 minutes (Unknown stream error: <stream:error><ack class="status" … type="media"/> in Evolution's log),
  delivering the same Status again each time. Messages arrived late but were not
  lost.
• Ingestion no longer stops after a few days. Every one of those reconnects
  left a Postgres connection pool open in Evolution, and after about a hundred
  of them its Postgres refused new connections: Evolution could no longer load
  the session, and nothing new was indexed.
• The session comes back on its own after a restart. Evolution now
  reconnects the paired number when its container starts
  (CONNECT_ON_STARTUP), instead of leaving WhatsApp disconnected until
  someone reconnected it from the panel.

The first two are bugs in Evolution Go 0.7.2
(evolution-go#211,
#185).
Their fix, evolution-go#190,
has not been released, so the stack now runs an unofficial build of 0.7.2 with
it applied: ghcr.io/brorlandi/whatsapp-mcp:evolution-0.7.2-patch.1, built by
this repository (see deploy/evolution). The
stack goes back to the official image once Evolution publishes the fix
(#6).

Updating

update.sh and the panel's Atualizar button bring the new compose file and the
new Evolution image. Evolution's container is recreated, which reconnects
WhatsApp once; the pairing is kept, so no QR code is needed.

• If you copied the compose file, take the new one. It changes the
  evolution-go image and adds CONNECT_ON_STARTUP: "true".
• If Evolution's Postgres already ran out of connections (pq: sorry, too many clients already in Evolution's log), recreating Evolution's container
  releases them.
• EVOLUTION_IMAGE selects another Evolution image; leave it unset.

v1.3.0-beta.1

The server version catches up with WhatsApp MCP Local 1.3.0: the same 40 MCP
tools as Local (plus send_contact and set_transcription_key), webhooks, and the
same control panel.

The version number jumps from 0.4 to 1.3 to match WhatsApp MCP Local: from
here on, the same version on both means the same tools and the same panel.
The -beta stays until this release has been run against a real WhatsApp
session; the schema can still change before 1.3.0.

New tools

• Reply, mention and draft. send_text_message quotes an earlier message
  (reply_to), mentions people in a group (mentions, written in the text as
  @Number) and, with dry_run, returns the draft and its recipient without
  sending. send_media_message gains reply_to, dry_run and stickers.
• Forward, mark as read, "typing…". forward_message resends a message
  marked as forwarded, media included; mark_chat_read clears a chat on the
  phone and sends the blue ticks; send_typing shows "typing…" or
  "recording audio…".
• Who is waiting for an answer. list_unanswered lists the chats whose
  latest message is from the other side, ignoring a closing "ok" or "obrigado",
  optionally with the groups that mentioned the account; list_unread lists
  what the phone shows as unread; list_mentions the messages that mention the
  account. mark_handled and snooze_chat take a chat off those lists until
  someone writes again — kept in the gateway, invisible to the other side.
• Counting and exporting. message_stats counts by chat, sender, day or
  month; get_message_context reads around one message; export_messages
  writes NDJSON and hands it over through a temporary link;
  get_chat_messages and search_messages take fields,
  max_content_chars and (reading) count_only.
• Groups. manage_group_participants, update_group,
  get_group_invite_link and leave_group. Removing someone, resetting the
  link and leaving ask for confirmation.
• health: one verdict and the checks behind it.
• Kept files. download_media keeps a copy of each file on the server's
  new data volume — it stays readable after WhatsApp discards it — and returns
  it as an image, an audio or a file block; files over 20 MiB come as a link.
  media_stats and purge_media measure and clear the volume.
• transcribe_audio gives Whisper the conversation as context and returns it
  with a review instruction; save_transcript keeps a correction's original
  as raw_text. Every message tool takes an optional chat_jid.
• Every tool carries MCP annotations (readOnlyHint, destructiveHint), and
  the server's instructions tell the assistant about webhooks.

Webhooks

• Configurações › Webhooks in the panel: a script receives every new
  message, reaction or read receipt as it arrives, signed with HMAC-SHA256.
  Add, test, turn off and delete each one, and see its last delivery; a
  webhook that stops answering is turned off by itself. A documentation page
  shows each delivery's JSON. The deliveries are the same as Local's, plus
  instance_id. See docs/webhooks.md.

The panel

• The panel now looks and works like Local's: Configurações behind the
  gear (MCP address, account, webhooks, transcription, updates, appearance,
  downloaded files, data), tabs Conectar MCP, WhatsApp, Status,
  Funções, Receitas and Ajuda, a step-by-step page per AI tool
  (Claude Desktop, Claude Code, ChatGPT / Codex, Cursor, another one) that
  finishes by itself when the tool connects, connections shown with each
  tool's logo, light, dark or system theme, and a Help tab with the common
  questions and examples. The old addresses redirect.

Ingestion

• The index reads quotes, mentions, forwards, reactions, polls and media file
  names, and applies edits and deletions to the message they change. Rows
  indexed before are filled in from their stored events on the first start.
• The gateway consumes READ_RECEIPT (the new receipt queue), which is how
  reading a chat on the phone clears it in list_unread.

Updating to this version

• update.sh brings the new docker-compose.yml: Evolution restarts with
  READ_RECEIPT in AMQP_GLOBAL_EVENTS, and the gateway gets the
  whatsapp_mcp_data volume (MEDIA_DIR, EXPORT_DIR) and INTERNAL_URL.
  Migrations 014 to 017 add the webhooks, settings, triage marks, chat
  state, message details and the transcript's raw text.
• Unread counts start from the moment of the update for chats the gateway has
  no read information about; a history sync or reading the chat on the phone
  fills it in.

Fixes

• Media WhatsApp has already discarded — its servers keep a file only for a
  while after it is sent — is now reported as that. download_media says the
  media expired and that only the sender resending it brings it back;
  link: true fetches the file before handing out a URL, so it fails there
  with the reason instead of giving a URL that answers a bare 502, and keeps
  the file for the link so the download is immediate. A large file that was
  not kept and has expired by the time it is fetched answers 410.
• The index's gap warnings sent the client to backfill_gap, a tool that
  does not exist. They now name sync_history with before.

Full Changelog: v0.4.0-beta.1...v1.3.0-beta.1

v0.4.0-beta.1

The panel can update the server it runs on, voice notes can be transcribed on
the user's own machine, and the stack fits a 1 GB server.

Updating from the panel

• The update notice now has an Atualizar button. It runs the same
  update.sh an operator would run by SSH — database dump first — and the
  panel follows its progress until the new version answers. The gateway never
  touches Docker: the button leaves a request in a directory shared with the
  host, and a systemd unit installed by install.sh (and by the next
  update.sh) carries it out. It accepts only a version GitHub has published.
• A Dokploy deployment can use the same agent in dokploy mode, which moves
  WHATSAPP_MCP_TAG through Dokploy's API and redeploys.
• The gateway remembers the newest version that ever ran against its
  database. Starting an older one — a redeployed stale pin, an update undone —
  raises a warning on every panel page, since migrations do not run backwards,
  with a button back to that version when the agent is there.
• The gateway container now runs as uid 10001, so the host can give it the
  shared directory. Migration 013_deployed_version.sql adds one table.

Local transcription

• The tool descriptions now tell a client that can run commands on the
  user's machine to transcribe locally first — mlx-whisper on Apple
  Silicon, faster-whisper or whisper.cpp on an NVIDIA GPU — and to use
  transcribe_audio (OpenAI) only when that is not possible. Local is free and
  the audio never leaves the machine.
• download_media takes link: true and answers with a temporary URL and a
  curl command instead of the base64 file, so a voice note does not have to
  pass through the conversation. The link needs no credential, is bound to one
  message and expires after ten minutes.
• save_transcript stores a transcript made elsewhere. get_chat_messages
  and search_messages now return every kept transcript in the message's
  transcript field, and search matches words spoken in voice notes.

Smaller machines, and AWS

• The stack runs on 1 GB of RAM. On a machine with less than 2 GB and no swap,
  the installer adds a 2 GB swap file, which is what gets a 1 GB plan through
  the image pulls and the first history sync. A 1 GB AWS Lightsail instance ran
  the whole stack at about 520 MB used, with nothing killed.
• docs/aws.md and docs/aws.pt-BR.md walk
  through Lightsail: the US$ 7 plan, why the price is the same in every
  region, pinning a static IP before installing, opening 443, the
  builder.aws.com projects whose region is fixed, and the AWS CLI commands an
  agent such as Claude Code can run while the user only signs in. The
  README's install prompt offers that route.

Fixes

• A fresh install failed at "starting the stack": quay.io/minio/minio
  started answering 401, after MinIO archived its community edition and
  stopped serving it anywhere. MinIO is gone from the stack instead of being
  pulled from somewhere else — it never held anything. Evolution only writes
  media to it with WEBHOOK_FILES on, which this stack leaves off, and no
  bucket was ever created; the gateway fetches media from WhatsApp through
  /message/downloadmedia. An existing instance loses the idle container on
  its next update, and the empty whatsapp-mcp_minio_data volume can be
  removed with docker volume rm whatsapp-mcp_minio_data. The MINIO_*
  lines left in its .env are ignored.
• Voice notes that arrived through a history sync could not be downloaded or
  transcribed ("the stored payload carries no media"). The media lookup only
  understood live events; a history sync stores whole conversations, and the
  message is now found among them by its id.
• A conversation WhatsApp moved to a LID (…@lid) was split in two: the
  history under the phone number, the newest messages under the LID. The
  gateway now records which LID belongs to which number — from the
  RecipientAlt/SenderAlt of live messages and the pnJID of history syncs
  — and list_chats shows one conversation, get_chat_messages reads both
  JIDs whichever one is asked for. Messages keep the JID WhatsApp used, so
  deleting, editing and reacting still address them correctly. Migration
  012_jid_aliases.sql pairs what is already indexed; LID chats whose number
  WhatsApp never revealed stay on their own.

v0.3.0-beta.2

Fixes

• transcribe_audio failed on every voice note with "Evolution returned no
  audio". Evolution Go returns the media as a data URI
  (data:audio/ogg; codecs=opus;base64,…) with an empty mimetype, and the
  tool was decoding it as bare base64. It now reads the format from the URI and
  falls back to Ogg/Opus, which is what WhatsApp voice notes are.

v0.3.0-beta.1

Voice notes can now be read. An instance that saves an OpenAI key gets its
WhatsApp audio transcribed on request.

Voice-note transcription

• transcribe_audio turns a WhatsApp voice note into text with OpenAI's
  Whisper. An AI client cannot hear the audio download_media returns, so the
  gateway does this one piece of model work itself — only with an OpenAI key the
  operator saves, and billed to that key.
• The key is saved from the panel's new Transcrição page or with the
  set_transcription_key tool. Both check it with OpenAI before saving, and
  neither shows it again: only its last four characters.
• The Transcrição page walks someone who has never used the OpenAI
  platform through getting a key — account, credit, key — with a button to
  each OpenAI page and what it will cost. Once a key is saved, it links to
  OpenAI's usage, credit, spending-limit and key pages instead.
• When there is no key, or OpenAI refuses it, transcribe_audio answers with
  the same steps and links, so the AI client can guide the user through setup
  instead of reporting an error.
• Every transcript is kept, so asking again for the same voice note costs
  nothing. whatsapp_status reports whether transcription is configured.
• The update adds two tables (migration 011_transcription.sql); nothing
  existing changes. The gateway needs outbound HTTPS to api.openai.com for
  transcription to work.

Publishing

• The latest image tag and the GitHub release marked Latest now follow the
  newest version, beta included. Pinning nothing used to leave an instance on
  0.1.0 — dozens of commits old, under the previous licence — while the beta
  was the version to run.

v0.2.0-beta.1

Full Changelog: v0.1.0...v0.2.0-beta.1

Full Changelog: v0.1.0...v0.2.0-beta.1

v0.1.0

Full Changelog: https://github.com/BrOrlandi/whatsapp-mcp/commits/v0.1.0

Full Changelog: https://github.com/BrOrlandi/whatsapp-mcp/commits/v0.1.0

Full Changelog: https://github.com/BrOrlandi/whatsapp-mcp/commits/v0.1.0