Instant Reply MCP
Connect AI agents to Instagram, WhatsApp Business, and Messenger inboxes, subject to workspace access and plan limits.
Hosted MCP Server
npx add-mcp 'https://api.instantreply.co/mcp'Installs into Claude Code, Codex, Cursor and more
Documentation
InstantReply MCP server for WhatsApp, Instagram, and Messenger AI agents
InstantReply MCP connects Claude, ChatGPT, Cursor, Codex, or any MCP-compatible agent to a real social inbox. The agent pairs a workspace through one browser link, answers questions with Barq, audits conversations, qualifies leads, validates WhatsApp templates, and operates approved workflows.
No API key is needed. Add the server, tell your agent to connect InstantReply, and open the link it gives you.
Set up in 60 seconds
1. Add the server to your agent (local package, nothing to install first):
| Client | How |
|---|---|
| Claude Code | claude mcp add instantreply -- npx -y @instantreply.co/mcp |
| Codex CLI | codex mcp add instantreply -- npx -y @instantreply.co/mcp |
| Claude Desktop | Settings, Developer, Edit config, paste the JSON below, restart |
| Cursor | Settings, MCP, Add server (or ~/.cursor/mcp.json), paste the JSON below |
| VS Code | .vscode/mcp.json, use the servers JSON below |
| Windsurf | ~/.codeium/windsurf/mcp_config.json, paste the JSON below |
| Anything else | Run the command npx -y @instantreply.co/mcp over stdio |
{
"mcpServers": {
"instantreply": {
"command": "npx",
"args": ["-y", "@instantreply.co/mcp"]
}
}
}
VS Code uses "servers" instead of "mcpServers" and needs "type": "stdio" next to command.
Remote (ChatGPT and other hosted clients) : add the connector URL https://api.instantreply.co/mcp. It uses OAuth 2.1 with PKCE; you approve access on an InstantReply consent page and never copy a key.
2. Tell your agent:
Connect InstantReply for my business.
3. What happens, and what the agent should say at each step
| Step | Agent action | Say to the user |
|---|---|---|
| 1 | Call get_connection_guide | "Instagram connects same-day. WhatsApp needs Meta Business Verification (usually 2-6 weeks). Which do you want first?" |
| 2 | Call start_setup (platform, optional persona: creator, business, developer, agency) | "Open this link, sign in or sign up, connect your account and approve the code ABCD-1234. I will wait." |
| 3 | Call check_setup until approved | "Waiting for your approval. Tell me when you have clicked approve." |
| 4 | Ask the user to reconnect or restart the server so all tools load | "Connected. Please restart this MCP server once." |
| 5 | Call list_channels | "These accounts are live: ... Nothing else is connected yet." |
| 6 | Optional: connect_channel for another Instagram account, Messenger page or WhatsApp number | "Open this second link to connect another account." |
| 7 | Teach the business with propose_knowledge_entry | "I drafted these facts about your business. Approve them in your dashboard or tell me to approve." |
Never ask the user to paste a secret into chat. The pairing key is stored locally at ~/.instantreply/config.json (file mode 0600). Set INSTANTREPLY_API_KEY for CI, Docker or shared machines.
Skills that ship with the package
The skills/ directory holds agent skills (SKILL.md with frontmatter) for the main jobs. Copy a folder into your agent's skills directory (for example ~/.claude/skills/) or point your agent at the folder.
| Skill | Job |
|---|---|
instantreply-setup | Pair, connect socials, verify what is live |
instantreply-inbox-triage | Find urgent, unanswered and high-intent conversations and draft replies |
instantreply-campaigns | Compliance-first outreach: opt-in, STOP, templates, the 24-hour window |
instantreply-brand-voice-and-kb | Teach the AI your business, voice and facts, with owner approval |
instantreply-add-integration | Connect your own system through a signed reply webhook, or add more channels |
Every skill references real tool names only; a test fails the build if one drifts.
Add your own integration
- Your backend answers customers:
set_channel_ai_providerwithprovider: custom_reply_webhook. InstantReply POSTs a signed request (X-InstantReply-Timestamp,X-InstantReply-Signature: sha256=HMAC-SHA256(secret, timestamp + "." + body)) to your HTTPS endpoint, which returns{"reply": "..."}. Replies still pass grounding checks. - More social accounts:
connect_channelreturns a link; the owner approves Meta OAuth in the browser. - Events pushed to your server: signed outbound webhooks are managed with the REST API (
/v1/webhooks) or the dashboard; there is no MCP tool for that yet.
What the agent can do
- Inbox work: list and inspect conversations, messages, contacts, channels, analytics, and usage.
- Comments: list, read, and publicly reply to Instagram/Facebook post comments; toggle AI comment auto-replies.
- Per-channel control: turn AI DM auto-replies on or off for one connected channel without touching the others.
- Barq assistance: ask product and workspace questions with
ask_barq; risky actions come back as approvals you confirm withdecide_approval. - Lead intelligence: audit recent conversations, find buying signals, segment contacts, draft follow-ups, and estimate campaign cost.
- WhatsApp operations: validate, create, submit, inspect, and delete templates; validate journeys; trigger individual or batch journeys; inspect delivery failures.
- Developer operations: inspect capabilities, onboarding requirements, and limits; troubleshoot delivery.
- Safety: state-changing tools are marked destructive so a careful client can ask for confirmation first.
Version 0.8.0 defines 87 workspace tools plus get_connection_guide (up to 88 tools once authenticated), 13 reusable prompts, an interactive workspace resource, and conversation/contact resource templates. Before pairing, only get_connection_guide, start_setup, and check_setup are available. Remote MCP plans omit ask_barq, propose_knowledge_entry, and propose_follow_up_rule; directory deployments also omit recommend_plan, set_channel_ai_provider, and decide_approval. Availability depends on plan, scopes, and connected channels. The exact tool schema is the source of truth.
Channel and Meta requirements
Instagram, Messenger, and WhatsApp availability depends on scopes, connected assets, and Meta approval state. WhatsApp Business API features can require Meta Business Verification, a connected WhatsApp Business Account, and approved templates. The agent should call get_connection_guide and inspect readiness instead of promising that every WhatsApp feature is immediately available.
InstantReply uses official Meta-connected workflows. Do not use the MCP server to send unsolicited messages, bypass consent, or evade Meta template and messaging-window rules. Honour STOP and unsubscribe requests immediately.
Environment variables
| Variable | Default | Purpose |
|---|---|---|
INSTANTREPLY_API_KEY | none | Skip browser pairing and authenticate directly |
INSTANTREPLY_API_URL | https://api.instantreply.co | Point the client at another environment |
Development
npm install
npm run build # compile dist/
npm test # run the MCP test suite
npm run dev # watch mode
Before publishing, run npm test, npm run build, npm pack --dry-run, and validate server.json with the official MCP Registry tooling. Keep package.json, package-lock.json, server.json, the README, and the vendored backend MCP copy on the same release version.
Links
- InstantReply MCP landing page
- API and developer documentation
- Remote MCP endpoint
- OAuth metadata
- Official MCP Registry
- npm package
License and hosted service
The public GitHub repository and npm client are MIT-licensed. This license covers the client only; the hosted InstantReply server and AI require an eligible plan. Free workspaces have read-only access. MCP Starter is $12/month ($120/year), MCP Growth is $29/month ($290/year), and MCP Pro is $89/month ($890/year). MCP plans use your own agent and model and include no InstantReply-hosted AI; choose an eligible Business or Developer plan for hosted AI. Yearly plans start with a 7-day free trial, card required, with no charge until day 8. Monthly plans have no trial and are charged immediately. Meta bills WhatsApp delivery fees separately.
Terminal
Run it in a terminal and it opens a launcher (Barq flies in, then a menu). When an AI tool starts it (piped, no terminal) it is the MCP server, exactly as before.
npx @instantreply.co/mcp # the launcher
npx @instantreply.co/mcp init # pair your account and add InstantReply to Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, Codex, Gemini CLI and Zed
npx @instantreply.co/mcp doctor # check Node, network, key, your AI tools, the server
npx @instantreply.co/mcp status | channels | tools [search] | call <tool> [json] | live | open | logout | uninstall
init --dry-run shows the exact change to each tool's config and writes nothing; configs are merged without touching your other entries or comments, a .bak is kept, and no secret is ever written into them (the key stays in ~/.instantreply). Every command takes --json. Animation is skipped when output is not a terminal, with NO_COLOR, CI or --no-animation. The daily version check can be turned off with INSTANTREPLY_NO_UPDATE_CHECK=1.
