ChatFlow MCP
Connect your AI agent to your WhatsApp CRM.
ChatFlow is a WhatsApp/Telegram/Instagram CRM for small
teams — a shared inbox, a drag-and-drop sales pipeline, AI auto-replies, and
broadcast campaigns. This is its hosted MCP (Model Context Protocol) server:
point Claude, Claude Code, Cursor, or any MCP-compatible agent at your workspace
and it can search contacts, read conversations, manage your pipeline, and — when
you allow it — send messages.
- Endpoint:
https://valuable-sockeye-14.eu-west-1.convex.site/mcp
- Transport: Streamable HTTP (stateless)
- Auth:
Authorization: Bearer <API key> — create one in ChatFlow →
Settings → AI agents (MCP)
Quick start
- In ChatFlow, open Settings → AI agents (MCP) (workspace owner) and click
Create API key. Copy it — it's shown once. Optionally give it an expiry.
- Add the server to your client:
Claude Code
claude mcp add --transport http chatflow \
https://valuable-sockeye-14.eu-west-1.convex.site/mcp \
--header "Authorization: Bearer cfk_YOUR_KEY"
Claude.ai / Claude Desktop — Settings → Connectors → Add custom connector →
paste the endpoint URL and set the Authorization: Bearer cfk_YOUR_KEY header.
Cursor — .cursor/mcp.json:
{
"mcpServers": {
"chatflow": {
"url": "https://valuable-sockeye-14.eu-west-1.convex.site/mcp",
"headers": { "Authorization": "Bearer cfk_YOUR_KEY" }
}
}
}
Full setup and troubleshooting: docs/setup.md.
| Tool | What it does |
|---|
search_contacts | Find contacts by name/phone substring |
list_conversations | Recent chats with previews, unread counts, CRM fields |
get_conversation | One chat: contact, notes, deal value, recent messages |
list_pipeline | Kanban columns with the leads in each stage |
move_conversation | Move a lead to a pipeline column |
update_lead | Set notes / deal value on a conversation |
create_contact | Create a WhatsApp contact + conversation from a phone number |
send_message | Send a text to the customer (opt-in — see below) |
get_usage | Subscription tier, limits, current monthly usage |
Read tools are annotated read-only; send_message is annotated as a
world-affecting write so well-behaved clients confirm before calling it.
Sending is opt-in
send_message messages a real customer, so it's off by default. The
workspace owner enables it with the "Allow agents to send messages" toggle
in Settings. While off, agents can read everything and draft replies, but every
send is rejected with a message explaining the toggle. When on, agent sends
count toward your monthly quota, appear in the inbox with the AI badge, pause
the auto-reply assistant on that chat (so two AIs never answer at once), and
support an idempotencyKey so a retry can't double-send.
Prompt-injection note: message text and notes an agent reads are written by
your customers. Keep the send toggle off unless you trust the agent driving it.
Links