Turbo Notify: WhatsApp for AI agents

Turbo Notify Agent Skills

Skills that put the Turbo Notify MCP server to work inside your AI agent: Claude Code, Claude Desktop and Cowork, Codex, the Antigravity CLI and the Gemini CLI, and any other agent that reads the Agent Skills format.

Skills

Skill	What it does
virtual-attendant	Serves a business's customers on WhatsApp: answers in the customer's language, remembers each customer, calls a person when it should, and takes commands from administrators and attendants on WhatsApp itself.

Customer guides (Portuguese and English): WhatsApp virtual attendant, Agent skills and AI agents, with one page per agent.

Install

Agent	Commands
Claude Code	claude plugin marketplace add turbo-notify/agent-skills then claude plugin install turbo-notify@turbo-notify
Claude Desktop and Cowork	Customize, Plugins, Add, Add marketplace, Add from a repository: turbo-notify/agent-skills; then Discover, turbo-notify, Add
Codex	codex plugin marketplace add turbo-notify/agent-skills then codex plugin add turbo-notify@turbo-notify
Antigravity	agy plugin install https://github.com/turbo-notify/agent-skills
Gemini CLI	gemini extensions install https://github.com/turbo-notify/agent-skills
Hermes Agent	hermes skills install turbo-notify/agent-skills/skills/virtual-attendant, then add the MCP server https://mcp.turbonotify.com in Hermes
OpenCode	npx skills add turbo-notify/agent-skills -a opencode (project folder .agents/skills/, or -g for ~/.config/opencode/skills/), then opencode mcp add --global --url https://mcp.turbonotify.com turbo-notify and opencode mcp auth turbo-notify to sign in. Guide: Turbo Notify in OpenCode
Cline	Add the MCP server https://mcp.turbonotify.com to cline_mcp_settings.json (steps for the editor and the CLI in llms-install.md)
Other agents that read the Agent Skills format (Cursor, Copilot and more)	npx skills add turbo-notify/agent-skills, then add the MCP server https://mcp.turbonotify.com in that agent's own settings

The unattended runner exists for Claude Code and Codex only. In the other agents the attendant runs in assisted mode: a session you keep open, where a person approves each draft.

The plugin rows above already declare the Turbo Notify MCP server (https://mcp.turbonotify.com), so there is nothing to add. Sign in with your Turbo Notify account the first time the agent uses it (Claude Code: /mcp; Codex: codex mcp login turbo-notify; Antigravity: Agent settings, Customizations, Authenticate next to the server), then ask it to "set up the Turbo Notify virtual attendant". No account yet? Create one at https://dashboard.turbonotify.com/auth/signup?locale=en. For anything that runs alone, use an API key instead (see the guide for your agent in skills/virtual-attendant/references/).

Update

The skill checks its version on every cycle and messages the administrators when a new one exists, with the command for the agent in use.

Agent	Command
Claude Code	claude plugin marketplace update turbo-notify then claude plugin update turbo-notify@turbo-notify
Claude Desktop and Cowork	The marketplace syncs from the repository; if a new version does not show up, remove the marketplace and add it again
Codex	codex plugin marketplace upgrade then codex plugin add turbo-notify@turbo-notify
Antigravity	Run the install command again
Gemini CLI	gemini extensions update turbo-notify

What it reaches

• The Turbo Notify MCP server at https://mcp.turbonotify.com, which the plugin declares in .mcp.json, mcp.json and mcp_config.json (one file for each family of agents; none holds a key or a header). You sign in with your account, or you connect your own with an API key. It runs with your account or API key permissions.
• One folder on your computer (by default ~/.turbo-notify/attendant) holding, for each customer, the conversation (kept 90 days by default, then deleted), short summaries of older parts of it, and a file with the facts that customer gave (name, preferences, what they asked for), plus what you taught. Turbo Notify gets no copy. /forget erases one customer's data, and the retention is a number you can change in memory-policy.json.
• The Turbo Notify API at https://api.turbonotify.com, called by the helper with your API key when the unattended runner or a console command sends a message.
• One network address besides Turbo Notify: raw.githubusercontent.com/turbo-notify/agent-skills/main/latest.json, read about once an hour (and again after 10 minutes if a check fails) for the version check. A failure changes nothing.
• Your other MCP servers only if you allow each one during setup.

What it runs and what it sends

• It runs the tn_assist helper with python3. The helper uses only the standard library and installs nothing.
• In the unattended runner (Claude Code and Codex), it starts your own claude or codex command once per customer, in a separate process with no tools, to write that customer's answer. The process gets that customer's messages, their file and the business knowledge that matches the message (the company, tone and hand-off files, plus up to three more picked by the words of the message), and nothing else.
• The text of a customer's message goes to the model of the agent you use, under your account with that agent. It goes nowhere else.
• The unattended runner reads your API key from the TURBO_NOTIFY_API_KEY variable you set for it and sends it only to api.turbonotify.com, as the sign-in of each request. It is never passed to the process that writes a customer's answer.
• Answers go out through the Turbo Notify API with your API key, to the customer who wrote. A check before every send stops another customer's number or name, non-public text and credentials.
• The first answer of each chat tells the customer that it comes from an AI. The line is added by the helper, not by the model.
• A customer who only says thanks, "ok" or sends a thumbs up gets no answer and is not handed to a person. The model has to ask for that, and the helper agrees only when the message is nothing but a thanks or an acknowledgement (no question mark, no number, a few words from a fixed list). Anything more is answered, or goes to a person.
• A customer goes to a person when the answer is not in what you taught (a price, stock, a deadline, a promise, a policy), when they ask for one or are upset, when the request is outside the business, when a payment problem, a refund or exchange, a complaint or a legal matter comes up, or when they ask the assistant to promise or confirm something in writing. A plain question about a rule you wrote, such as the exchange window, is answered.
• If the model's answer cannot be read, the helper asks again once with a line naming what was wrong, and a second failure in a row sends the customer a message saying a person will continue.

Layout

.claude-plugin/        plugin and marketplace manifests (Claude Code, Codex)
.codex-plugin/         the manifest Codex reads for the name, logo and listing text
plugin.json            manifest for Antigravity
gemini-extension.json  manifest for the Gemini CLI (the connection to the server)
.mcp.json              the Turbo Notify MCP server, as Claude Code, Claude Desktop, Cowork and Codex read it
mcp.json               the same server, as the Agent Plugins format reads it (Copilot, Cursor)
mcp_config.json        the same server, as Antigravity reads it
assets/icon.png        the plugin icon
agents/                the tool-less agent that answers a customer
skills/virtual-attendant/
  SKILL.md             the entry point the agent reads
  references/          one short file per topic (setup, tick, console, memory, privacy, ...)
  assets/templates/    interview templates (pt, en) and the memory policy
  scripts/tn_assist/   the helper: routing, memory, egress check, runner (Python 3.9+, no dependencies)
tools/release.py       version consistency check and sync
tests/                 unit, property and security tests (Poetry, development only)
evals/attendant/       live evaluation of a customer's turn on a real model, with a made-up shop

The helper holds the guarantees the model cannot be trusted with: who is who is decided by code from the registered numbers, a customer's turn runs without tools, every reply passes an egress check before it can leave, and the data never leaves the workspace folder.

Development

poetry install
poetry run pytest
poetry run ruff check . && poetry run ruff format --check .
poetry run mypy
python3 tools/release.py check

The live evaluation needs your own claude command, logged in. It asks 40 questions that the knowledge cannot answer and 11 that it can, and plays seven conversations (some end with a thank you, which is owed no answer). It checks every reply for a price, a date, a time, a quantity, a discount or a promise the knowledge does not state, for the language of the reply, and that a thank you is left alone while an "ok" that is really a question is answered. Two last checks cut the model's answer in half to see what the runner does with an answer it cannot read, and play an "ok" after an offer ("quer que eu chame alguém?") to see that it is answered, not left alone. Nothing in it is real data.

poetry run python -m evals.attendant run                # writes evals/attendant/results/<date>-<label>/
poetry run python -m evals.attendant run --set refusal  # only the check on an unreadable answer
poetry run python -m evals.attendant run --set offers   # only the check on an ok that accepts an offer
TN_EVALS_LIVE=1 poetry run pytest evals                 # the same, as tests (skipped otherwise)

Releases bump VERSION, latest.json, the five manifests, the helper and pyproject versions and CHANGELOG.md together (tools/release.py sync), then move the v1 tag, which stays available for anyone who wants to pin to 1.x (...agent-skills#v1, or --ref v1 in Codex).

License

MIT. See LICENSE.