TinkerClaw WhatsApp

You put 5 agents in a WhatsApp group. They all respond at once. Your API bill does a backflip. Protocol v2 fixes that — congestion control, conversation lifecycle, and budget-aware scheduling. Agents that know when to talk, when to shut up, and when to burn unused tokens before reset. Built for the TinkerClaw fork — github.com/globalcaos/tinkerclaw. Also ships four opt-in maintenance scripts that read your WhatsApp session credentials, enumerate group contacts, and patch your OpenClaw source tree — each refuses to run without an explicit --yes. See Permissions, Data Flow & Consent.

Oscar Serra@globalcaos

Install

openclaw skills install @globalcaos/whatsapp-ultimate

WhatsApp Ultimate

One of dozens of skills and plugins in TinkerClaw — a self-improving OpenClaw fork that's been running 24/7 for months.

You put 5 agents in a WhatsApp group. They all respond at once. Your API bill does a backflip.

Protocol v2 fixes that. Your agents learn group etiquette: when to talk, when to shut up, and when to spend the tokens they'd otherwise waste before the limit resets.

It adds congestion control so a busy group doesn't trigger a stampede of replies, a conversation lifecycle so a thread knows when it's actually over, and budget-aware scheduling so the cheap chatter waits and the work that matters gets the headroom. Multi-agent WhatsApp groups that stay calm, stay on-topic, and stay inside budget.

Part of TinkerClaw — real-time token tracking, self-improving crons, persistent cognitive memory. This is one piece of that stack; the repo has dozens more.

👉 https://github.com/globalcaos/tinkerclaw

Clone it. Fork it. Break it. Make it yours.

Everything you can do in WhatsApp, your AI agent can do too. This skill documents all WhatsApp capabilities available through OpenClaw's native channel integration. No external Docker services, no CLI wrappers — direct WhatsApp Web protocol via Baileys.

- OpenClaw with WhatsApp channel configured - WhatsApp account linked via QR code (`openclaw whatsapp login`)

Permissions, Data Flow & Consent

Short version: most of this skill is documentation and needs no permissions at all. The action reference below describes things OpenClaw's own WhatsApp channel already does — this package does not implement them and cannot do them on its own. What this package does ship is four scripts in scripts/, and those are genuinely powerful. They are all off by default: each one prints exactly what it is about to touch and then exits unless you pass --yes.

Longer version, because you should not have to take that on trust.

What it needs, and why.

Capability	Why	Scope	On by default?
Read WhatsApp credentials	The two Baileys scripts act as your linked device	`~/.openclaw/credentials/whatsapp/default`, or `--auth-dir` / `WA_AUTH_DIR`	No — `--yes` required
Network	Baileys → WhatsApp Web servers, as your account	WhatsApp only. No telemetry, no analytics, no third-party endpoint	No — `--yes` required
File write (export)	Contact/group inventory	`~/.openclaw/workspace/bank/whatsapp-contacts-full.json`, or `--save=PATH`; written mode `0600`	No — requires `--save`
File write (source)	The two `apply-*.sh` scripts edit your OpenClaw checkout in place	`<openclaw-src>/src/**`; `.bak` taken first, `--revert` restores	No — `--yes` required
Create groups / message	`wa-create-group.ts` creates a real, visible group	Only the group name + numbers you pass	No — `--yes` required
Read LID → phone mappings	Turn opaque WhatsApp LIDs into real phone numbers	Auth dir; only with `--resolve-lids`	No — masked to last-4 otherwise
Env read	`WA_AUTH_DIR`, `OPENCLAW_SRC`	Path overrides only	—
Secrets / tokens / API keys	None. Nothing reads an API key, and no token, key or secret is ever transmitted	—	—

The privacy-affecting things, stated plainly.

Three of these scripts touch data about other people, who have not agreed to any of it. That is a real cost and the docs should say so rather than bury it:

Contact enumeration (wa-fetch-contacts.ts) builds a map of every group you are in and everyone in them. By default phone numbers are masked to the last 4 digits and nothing is written to disk — you get an in-memory summary. --resolve-lids un-masks them; --save persists them at mode 0600. Both are deliberate, separate opt-ins.
Message retention (apply-history-fix.sh) makes OpenClaw store every inbound message — text, sender, timestamp — in a local SQLite DB, with no retention limit and no encryption beyond your filesystem's. Applying it is a decision about other people's messages. In some jurisdictions it is also a legal obligation you are taking on. Off unless you apply it.
Identity metadata (apply-model-prefix.sh) can print which auth profile served a reply into the outgoing message, where everyone in the chat sees it. It reads the profile name, never the token behind it.

Turning it off. Nothing here is on until you turn it on, and everything is reversible:

bash

# The scripts do nothing without --yes. That is the off switch: don't pass it.
./scripts/apply-history-fix.sh --revert     # un-patch message retention
./scripts/apply-model-prefix.sh --revert    # un-patch the auth-mode prefix
rm ~/.openclaw/workspace/bank/whatsapp-contacts-full.json   # delete the export

To stop the auth-mode prefix without reverting anything, just remove {authMode} / {authProfile} from responsePrefix in your config. To stop history capture without reverting, delete the history DB — but note that new messages keep arriving until you revert.

Read them before you run them. All four scripts are short and are meant to be read end to end. That is the whole security model.

Capabilities Overview

Category	Features
Messaging	Text, media, polls, stickers, voice notes, GIFs
Interactions	Reactions, replies/quotes, edit, unsend
Groups	Create, rename, icon, description, participants, admin, invite links
History	Local-DB search only (see caveat below). vCard contact extraction

Total: 22 distinct actions.

These 22 actions belong to OpenClaw's WhatsApp channel, not to this package. This skill documents them; installing it does not add them, and uninstalling it does not remove them. The code this package actually ships is the four scripts in scripts/ — documented below, all consent-gated.

Messaging

Send Text

text

message action=send channel=whatsapp to="+34612345678" message="Hello!"

Send Media (Image/Video/Document)

text

message action=send channel=whatsapp to="+34612345678" message="Check this out" filePath=/path/to/image.jpg

Supported: JPG, PNG, GIF, MP4, PDF, DOC, etc.

Send Poll

text

message action=poll channel=whatsapp to="+34612345678" pollQuestion="What time?" pollOption=["3pm", "4pm", "5pm"]

Send Sticker

text

message action=sticker channel=whatsapp to="+34612345678" filePath=/path/to/sticker.webp

Must be WebP format, ideally 512x512.

Send Voice Note

text

message action=send channel=whatsapp to="+34612345678" filePath=/path/to/audio.ogg asVoice=true

Use OGG/Opus format for voice notes — MP3 may not play correctly.

Send GIF

text

message action=send channel=whatsapp to="+34612345678" filePath=/path/to/animation.mp4 gifPlayback=true

Convert GIF to MP4 first (WhatsApp requires this):

bash

ffmpeg -i input.gif -movflags faststart -pix_fmt yuv420p -vf "scale=trunc(iw/2)*2:trunc(ih/2)*2" output.mp4 -y

Interactions

Add Reaction

text

message action=react channel=whatsapp chatJid="34612345678@s.whatsapp.net" messageId="ABC123" emoji="🚀"

Remove Reaction

text

message action=react channel=whatsapp chatJid="34612345678@s.whatsapp.net" messageId="ABC123" remove=true

Reply/Quote Message

text

message action=reply channel=whatsapp to="34612345678@s.whatsapp.net" replyTo="QUOTED_MSG_ID" message="Replying to this!"

Edit Message (Own Messages Only)

text

message action=edit channel=whatsapp chatJid="34612345678@s.whatsapp.net" messageId="ABC123" message="Updated text"

Unsend/Delete Message

text

message action=unsend channel=whatsapp chatJid="34612345678@s.whatsapp.net" messageId="ABC123"

Irreversible and visible. Deletion-for-everyone cannot be undone, and WhatsApp leaves a "This message was deleted" tombstone in the chat. Confirm with the human before an agent unsends anything it did not send itself.

Group Management

Create Group

text

message action=group-create channel=whatsapp name="Project Team" participants=["+34612345678", "+34687654321"]

Rename Group

text

message action=renameGroup channel=whatsapp groupId="123456789@g.us" name="New Name"

Set Group Icon

text

message action=setGroupIcon channel=whatsapp groupId="123456789@g.us" filePath=/path/to/icon.jpg

Set Group Description

text

message action=setGroupDescription channel=whatsapp groupJid="123456789@g.us" description="Team chat for Q1 project"

Add Participant

text

message action=addParticipant channel=whatsapp groupId="123456789@g.us" participant="+34612345678"

Remove Participant

text

message action=removeParticipant channel=whatsapp groupId="123456789@g.us" participant="+34612345678"

Affects a real person and requires admin. Removal is visible to the whole group and you cannot silently undo it. Same for promoteParticipant / demoteParticipant. Get explicit human confirmation before an agent changes a group's membership or admin list.

Promote to Admin

text

message action=promoteParticipant channel=whatsapp groupJid="123456789@g.us" participants=["+34612345678"]

Demote from Admin

text

message action=demoteParticipant channel=whatsapp groupJid="123456789@g.us" participants=["+34612345678"]

Leave Group

text

message action=leaveGroup channel=whatsapp groupId="123456789@g.us"

Get Invite Link

text

message action=getInviteCode channel=whatsapp groupJid="123456789@g.us"

Returns: https://chat.whatsapp.com/XXXXX

Revoke Invite Link

text

message action=revokeInviteCode channel=whatsapp groupJid="123456789@g.us"

Both directions are sensitive. getInviteCode returns a link that lets anyone holding it join the group — treat it as a credential and do not paste it into logs or other chats. revokeInviteCode instantly breaks every copy of the old link that is already circulating.

Get Group Info

text

message action=getGroupInfo channel=whatsapp groupJid="123456789@g.us"

Returns: name, description, participants, admins, creation date.

JID Formats

WhatsApp uses JIDs (Jabber IDs) internally:

Type	Format	Example
Individual	`<number>@s.whatsapp.net`	`34612345678@s.whatsapp.net`
Group	`<id>@g.us`	`123456789012345678@g.us`

When using to= with phone numbers, OpenClaw auto-converts to JID format.

History is a local-DB search. Live fetch does not work.

Your agent cannot pull old messages back off your phone on demand. Anything that claims otherwise is wrong. What actually exists:

Search of the local whatsapp-history.db — whatever your gateway captured while it was running, including media captions and link previews, not just plain text.
On-demand historySync does not land. The phone answers the request and the payloads are discarded before they reach storage. Treat a live channel plus an empty result as a gap in what was captured, not as a broken skill.
Any period your gateway was down is simply missing, and no amount of retrying will fill it in. Check your own downtime before assuming a bug.
The backfill that does work: use Export chat on your phone and drop the file into your workspace. Re-pairing via a fresh QR may bootstrap history, but it drops your existing linked session — that is a deliberate decision, not something an agent should do for you.

The history DB is off unless you turn it on. Capturing every inbound message is a decision about other people's data — see Permissions, Data Flow & Consent and scripts/apply-history-fix.sh.

Bundled Scripts (`scripts/`)

This is the code the package actually ships. All four are off by default — each prints what it will touch and exits unless you pass --yes. Read Permissions, Data Flow & Consent first.

The two .ts scripts drive Baileys directly and need Node plus @whiskeysockets/baileys and pino resolvable — run them from inside your OpenClaw checkout, which already has both. The two .sh scripts need bash and python3.

`wa-fetch-contacts.ts` — inventory your groups and their members

bash

npx tsx scripts/wa-fetch-contacts.ts                    # prints what it would do, then refuses
npx tsx scripts/wa-fetch-contacts.ts --yes              # in-memory only, numbers masked to last-4
npx tsx scripts/wa-fetch-contacts.ts --yes --save       # persist to bank/whatsapp-contacts-full.json (0600)
npx tsx scripts/wa-fetch-contacts.ts --yes --resolve-lids   # un-mask: full phone numbers

Flag	Effect
`--yes`	Required. Without it the script exits 1 and touches nothing
`--save[=PATH]`	Opt in to writing the dataset to disk, mode `0600`. Default: nothing is written
`--resolve-lids`	Opt in to reading LID→phone mappings and emitting full numbers. Default: last-4 masking
`--auth-dir=PATH`	Use a different session directory (also `WA_AUTH_DIR`)

This builds a dataset about people who did not consent to it. Delete the file when done.

`wa-create-group.ts` — create a group

bash

npx tsx scripts/wa-create-group.ts --yes "Project Team" "+34612345678" "+34687654321"

Reads your session credentials, connects as your account, and creates a real group that every participant sees appear on their phone. Not undoable from here. Only the group name and the numbers you pass are sent to WhatsApp. --auth-dir=PATH / WA_AUTH_DIR override the credential path.

`apply-history-fix.sh` — turn on full inbound message retention

bash

./scripts/apply-history-fix.sh --yes      # patch monitor.ts (a .bak is taken first)
./scripts/apply-history-fix.sh --revert   # restore from the .bak

Edits src/web/inbound/monitor.ts in your OpenClaw checkout so every inbound message — text, sender JID, push name, timestamp — is written to the local history DB, including self-chat messages Baileys otherwise drops. No retention limit, no redaction, no encryption beyond your filesystem's. This is a decision about other people's messages; in some jurisdictions it is also a notice obligation you are taking on. After applying: rebuild and restart your gateway.

`apply-model-prefix.sh` — show model + auth mode in outgoing messages

bash

./scripts/apply-model-prefix.sh --yes      # patch 4 files (.bak taken for each)
./scripts/apply-model-prefix.sh --revert   # restore from the .bak files

Adds {authMode} / {authProfile} to responsePrefix, e.g. "🤖({model}|{authMode})". It reads the auth profile's name — never a token, key or secret — but that name then appears in your outgoing messages where everyone in the chat can see it. To stop the disclosure without reverting, drop those variables from responsePrefix.

Tips

Voice Notes

Use OGG/Opus format:

bash

ffmpeg -i input.wav -c:a libopus -b:a 64k output.ogg

Stickers

Convert images to WebP stickers:

bash

ffmpeg -i input.png -vf "scale=512:512:force_original_aspect_ratio=decrease,pad=512:512:(ow-iw)/2:(oh-ih)/2:color=0x00000000" output.webp

Rate Limits

WhatsApp has anti-spam measures. Avoid:

Bulk messaging to many contacts
Rapid-fire messages
Messages to contacts who haven't messaged you first

Message IDs

To react/edit/unsend, you need the message ID. Incoming messages include this in the event payload. For your own sent messages, the send response includes the ID.

Comparison with Other Skills

Feature	whatsapp-ultimate	wacli	whatsapp-automation	gif-whatsapp
Native integration	yes	no (CLI)	no (Docker)	N/A
Send text	yes	yes	no	no
Send media	yes	yes	no	no
Polls	yes	no	no	no
Stickers	yes	no	no	no
Voice notes	yes	no	no	no
GIFs	yes	no	no	yes
Reactions	yes	no	no	no
Reply/Quote	yes	no	no	no
Edit	yes	no	no	no
Unsend	yes	no	no	no
Group create	yes	no	no	no
Group management	yes (full)	no	no	no
Receive messages	yes	yes	yes	no
Two-way chat	yes	no	no	no
External deps	None for the 22 actions; Node + Baileys for the bundled scripts	Go binary	Docker + WAHA	ffmpeg

Protocol v2: Multi-Agent Discussions

<why_this_matters> If you put multiple AI agents in one WhatsApp group, the naive default is everyone responds to everything. Five agents replying to one message means 5x the API spend per turn, plus echo loops where agents agree with each other forever. Protocol v2 introduces congestion control, conversation lifecycle, and budget-aware scheduling so agents know when to talk, when to stay quiet, and when to wrap up. </why_this_matters>

Agent Identity

Each agent gets its own personality, icon, and (optionally) model:

yaml

channels:
  whatsapp:
    agentIcon: "🤖"          # single-agent icon prefix
    turnEndMarker: "⚡"       # end-of-turn marker in 1:1 chats
    multiAgent:
      mainAgentId: "jarvis"
      agents:
        jarvis:
          id: "jarvis"
          name: "Jarvis"
          icon: "🤖"
        luna:
          id: "luna"
          name: "Luna"
          icon: "🌙"
          model: "sonnet"
        rex:
          id: "rex"
          name: "Rex"
          icon: "🦖"
          model: "haiku"

Agent personalities live in the workspace:

text

workspace/
├── SOUL.md                  # main agent
├── agents/
│   ├── luna/SOUL.md         # Luna's personality
│   └── rex/SOUL.md          # Rex's personality

Intra-Agent Chats

yaml

      intraAgentChats:
        brainstorm:
          chatId: "123456789012345678@g.us"   # your own group's JID
          participants: ["jarvis", "luna", "rex"]
          owner: "your-owner-id"
          mode: "broadcast"        # broadcast | addressed | round-robin

Routing modes:

broadcast — all agents respond (with congestion control)
addressed — only respond when mentioned by name ("Luna, what do you think?")
round-robin — structured turn-taking

Congestion Control (Exponential Courtesy Protocol)

Prevents N agents from all responding simultaneously:

yaml

      congestion:
        enabled: true
        baseDelayFactor: 150     # ms × agentCount² base delay
        maxDelay: 30000          # 30s cap
        backpressureThreshold: 1.5  # slow down over-talkers
        windowMs: 60000          # 60s sliding window

How it works:

Base delay scales quadratically with agent count (2 agents ≈ 600ms, 5 agents ≈ 3750ms)
Random jitter prevents synchronization
Agents talking more than their fair share get 2× delay penalty
If another agent posts during your wait, restart the timer (yield-on-collision)

Conversation Lifecycle

Agents detect when discussions go stale and know when to wrap up:

yaml

      lifecycle:
        stalenessWindow: 5        # compare last N messages
        stalenessThreshold: 0.85  # cosine similarity trigger
        maxTurnsPerObjective: 30  # hard cap
        autoClose: true

Features:

Staleness detection — cosine similarity of message embeddings detects circular discussions
Agreement loop detection — catches "I agree" / "Good point" / "Exactly" loops
Topic steering — one agent claims pivot role to redirect conversation
Objective tracking — set goals, track completion, auto-close with summary
Closure protocol — propose → ack → converge (all agents must agree)

Budget-Aware Scheduling

Adjusts conversation depth based on API usage and reset timing:

yaml

      budget:
        provider: "anthropic"
        windowDays: 7
        burnModeEnabled: true
        burnTriggerHours: 24     # hours before reset
        burnUsageThreshold: 0.20 # usage below 20%

Four modes:

Mode	When	Congestion	Staleness	Max Turns	Tangents
Conservative	>85% used	2× slower	0.80	½	No
Moderate	60-85%	Normal	0.85	Normal	No
Aggressive	<60%	0.7× faster	0.85	Normal	Yes
Burn	<20% used, <24h to reset	0.3× faster	0.95	2×	Encouraged

Burn mode philosophy: unused tokens expire at reset. Better to have emergent agent-agent discussions than waste the budget.

DM Trigger Prefix

Protocol v2 extends triggerPrefix to DMs (previously groups only):

Owner — always bypasses triggerPrefix
Authorized contacts — must start message with prefix (e.g., "Jarvis, help me with...")
Intra-agent chats — bypass triggerPrefix entirely

Turn-End Marker

In 1:1 chats (selfChat or owner-only DM), append a visual marker to signal turn completion:

yaml

channels:
  whatsapp:
    turnEndMarker: "⚡"

4.1.0

Added: "Permissions, Data Flow & Consent" — full capability table, what is read, what is written, where it goes
Added: Consent gate on all four bundled scripts — each prints what it will touch and exits unless you pass --yes
Added: --revert off-switch and automatic .bak for both source-patching scripts
Changed: wa-fetch-contacts.ts is now ephemeral by default — nothing is written to disk without --save, and saved files are mode 0600
Changed: wa-fetch-contacts.ts masks phone numbers to the last 4 digits; LID→phone resolution is opt-in via --resolve-lids
Added: --auth-dir / WA_AUTH_DIR override so credentials are not hard-coded to one path
Added: Documentation for all four bundled scripts, which were previously undisclosed
Added: Safety notes on unsend, participant/admin changes, and invite links
Fixed: "External deps: None" now distinguishes the channel actions from the bundled scripts
Fixed: Removed operator-specific content (real group JID, owner name, internal notes) from the history section and config examples

4.0.0

Protocol v2: Multi-agent discussions with configurable routing (broadcast/addressed/round-robin)
Added: Congestion control — Exponential Courtesy Protocol prevents message explosion in multi-agent chats
Added: Conversation lifecycle — staleness detection, agreement loop detection, topic steering, objective tracking, closure protocol
Added: Budget-aware scheduling — four spending modes including burn mode for pre-reset token usage
Added: Agent identity system — per-agent SOUL.md, icons, names, model overrides
Added: DM triggerPrefix gating — non-owner contacts must use prefix in DMs
Added: Turn-end marker (⚡) for 1:1 chats
Added: agentIcon config for outbound message prefixing

3.7.0

Added: vCard phone number extraction — contact messages now return structured vcard field with names and phone numbers
Added: contactsArrayMessage support — multi-contact shares are now parsed
Improved: New contact messages store phone numbers in text_content for full-text search (e.g. search by phone number)
Improved: raw_json now included in search results for contact-type messages, enabling vCard extraction from historical data

3.4.0

Fixed: Chat search now resolves LID/JID aliases — searching by chat name finds messages across both @lid and @s.whatsapp.net JID formats
Added: resolveChatJids() cross-references chats, contacts, and messages tables to discover all JID aliases for a given chat filter
Improved: Search falls back to original LIKE behaviour if no JIDs resolve, so no regressions

3.0.0

text

Your Agent
    ↓
OpenClaw message tool
    ↓
WhatsApp Channel Plugin
    ↓
Baileys (WhatsApp Web Protocol)
    ↓
WhatsApp Servers

No external services. No Docker. No CLI tools. Direct protocol integration.

Included Files

File	Purpose	Runs anything?
`SKILL.md`	This page — the action reference and the protocol v2 config	No
`description.md`	ClawHub listing copy	No
`scripts/wa-fetch-contacts.ts`	Group/contact inventory. Consent-gated, masked + ephemeral by default	Yes — `--yes`
`scripts/wa-create-group.ts`	Create a group via Baileys. Consent-gated	Yes — `--yes`
`scripts/apply-history-fix.sh`	Patch OpenClaw for full inbound retention. Consent-gated, `--revert`	Yes — `--yes`
`scripts/apply-model-prefix.sh`	Patch OpenClaw for `{authMode}` prefix. Consent-gated, `--revert`	Yes — `--yes`

There is no bin/, no installer, no background process and no telemetry. Everything the documentation above describes is either in this package or is a documented action of OpenClaw's own WhatsApp channel — and the table in Permissions, Data Flow & Consent says which is which. If you find a claim here that the code does not do, that is a bug — open an issue on the repo.

Pairs Well With

smart-model-router — auto-select the right model per agent role (creative → Sonnet, analyst → Haiku, devil's advocate → GPT)
agent-superpowers — verification iron law and three-agent review for when your multi-agent discussions produce code
subagent-overseer — monitor agent sessions without burning tokens on polling loops

https://github.com/globalcaos/tinkerclaw

Clone it. Fork it. Break it. Make it yours.

License

MIT — Part of OpenClaw

TinkerClaw WhatsApp

Install

WhatsApp Ultimate

Permissions, Data Flow & Consent

Capabilities Overview

Messaging

Send Text

Send Media (Image/Video/Document)

Send Poll

Send Sticker

Send Voice Note

Send GIF

Interactions

Add Reaction

Remove Reaction

Reply/Quote Message

Edit Message (Own Messages Only)

Unsend/Delete Message

Group Management

Create Group

Rename Group

Set Group Icon

Set Group Description

Add Participant

Remove Participant

Promote to Admin

Demote from Admin

Leave Group

Get Invite Link

Revoke Invite Link

Get Group Info

JID Formats

History is a local-DB search. Live fetch does not work.

Bundled Scripts (scripts/)

wa-fetch-contacts.ts — inventory your groups and their members

wa-create-group.ts — create a group

apply-history-fix.sh — turn on full inbound message retention

apply-model-prefix.sh — show model + auth mode in outgoing messages

Tips

Voice Notes

Stickers

Rate Limits

Message IDs

Comparison with Other Skills

Protocol v2: Multi-Agent Discussions

Agent Identity

Intra-Agent Chats

Congestion Control (Exponential Courtesy Protocol)

Conversation Lifecycle

Budget-Aware Scheduling

DM Trigger Prefix

Turn-End Marker

4.1.0

4.0.0

3.7.0

3.4.0

3.0.0

Included Files

Pairs Well With

License

Links

Related skills

Bundled Scripts (`scripts/`)

`wa-fetch-contacts.ts` — inventory your groups and their members

`wa-create-group.ts` — create a group

`apply-history-fix.sh` — turn on full inbound message retention

`apply-model-prefix.sh` — show model + auth mode in outgoing messages