# Connectors: your backend and your helpdesk

> Capabilities map to your endpoints; hand-offs land in the helpdesk you already run: Stripe, Zendesk, Slack, or a webhook.

Canonical: https://peeve.ai/product/connectors/

Capabilities map to your own endpoints (defined in peeve.config.json, reviewable in a PR). Hand-offs land in the helpdesk you already run: Zendesk, Intercom, Slack, HubSpot or a webhook. The model requests a capability by name; your execution service makes the call with your credentials, which the model never sees.

## Does real backend work: custom connectors
Peeve doesn't just answer from a help center. Connect your own tools two ways and, mid-conversation, the assistant reads live data (an order, an invoice, an account status) and takes real actions for your customers. Minutes, not weeks. No SDK, no integration project.

### Custom API Connector (Scale and up)
Turn your existing REST API into something the assistant can use, built automatically from your API docs.
1. Paste your API docs URL.
2. Peeve reads the docs and builds the connector: base URL, auth, and every endpoint, each marked read (safe lookup) or write (changes something).
3. Add your API credentials (encrypted, per workspace).
4. The assistant can now use your API in conversations.
Read endpoints are called live to answer with current data; write endpoints (a refund, a record update) always run through a confirmation step and a workspace allow-list.

### Custom MCP Server (Growth and up)
Connect an MCP server your team already runs (MCP = Model Context Protocol, the open standard for exposing tools to AI). An advanced option. Most customers never need it.
1. Paste your MCP server URL (and an access token if it needs one).
2. Peeve does a live handshake and discovers the tools your server exposes.
3. You choose which of those tools the assistant is allowed to use.
When the page alone can't answer, the assistant calls your MCP tools to fetch data or perform an action; anything that changes state runs through the same confirm guardrail.

Guardrails: reads are automatic and every write requires confirmation; you pick exactly which tools and endpoints are enabled; credentials are encrypted at rest per workspace and never shown again.

## The connector catalog
Connectors come in three shapes:
- **Outbound** — Peeve acts on a system you run: deliver a hand-off, book a meeting, read a plan, push a lead. Your credentials, encrypted per workspace; the model never sees them.
- **Migrate** — A one-time import of your existing help content or bot training into the Brain, so Peeve starts from what you already wrote. No re-authoring.
- **Custom (MCP)** — Point Peeve at your own MCP server and it connects as a client, exposing your tools the same way Claude uses a custom MCP.
- **Custom (API)** — Point Peeve at your REST API docs and it builds the connector for you: base URL, auth and every endpoint, each marked read or write.

### Support & messaging
Hand-offs and two-way conversations land in the tools you already run.

- **Email** — Send hand-offs from your own mailbox: Gmail, Outlook, or any SMTP.
- **Slack** — Post hand-offs and notifications into a channel.
- **Zendesk** — Create or append a ticket from a hand-off.
- **WhatsApp** — Two-way chat hand-off.
- **Messenger** — Two-way chat hand-off.
- **Instagram** — Two-way DM hand-off.

### Scheduling
Book meetings straight from the cursor.

- **Cal.com** — Surface availability and book an event type.
- **Calendly** — Share a scheduling link and book.

### Payments & subscriptions
Read a customer's plan and act on it.

- **Stripe** — Read a customer's plan and lifecycle; generate a checkout or portal link.

### CRM & leads
Read who someone is, and push the leads Peeve captures.

- **HubSpot** — Read a contact's plan and lifecycle; push captured leads.
- **Salesforce** — Read and write leads and contacts.

### Migrate your knowledge → the Brain
A one-time import of your existing help content or bot training, so Peeve starts from what you've already written.

- **Intercom / Fin** — Help-center articles and Fin content.
- **Chatbase** — Bot sources and training data.
- **Tidio / Lyro** — Knowledge base and Lyro training.
- **Crisp** — Helpdesk articles.
- **Freshchat / Freshdesk** — Solutions and articles.
- **Help Scout** — Docs articles.
- **Ada** — Bot answers.
- **Botpress / Voiceflow / Landbot** — Flows and knowledge base.
- **File import** — CSV, JSON, or Markdown export.

### AI agents (agent channel)
Your users point their own AI agent at a scoped Peeve endpoint and get your capabilities as typed tools, the same permissions they have, never more.

- **Claude** — Add your Peeve endpoint as a custom MCP connector.
- **OpenAI** — Connect via MCP and call your tools with stated effects.
- **Hermes** — Connect the Hermes agent over MCP to your Peeve endpoint.
- **OpenClaw** — Connect the OpenClaw agent over MCP to your Peeve endpoint.
- **Any MCP client** — Any agent that speaks MCP connects and gets your capabilities as typed tools.

### Custom backend (MCP & API)
Bring your own tools. Peeve looks up live data and takes action in your systems, with a confirmation step on anything that changes.

- **Custom MCP server** — Peeve initializes, lists, and calls your MCP tools, like Claude's custom MCP. Growth and up.
- **Custom API connector** — Built automatically from your REST API docs: reads live, writes with confirmation. Scale and up.


## More questions

### Does Peeve integrate with Zendesk and Stripe?

Yes. Hand-offs land in Zendesk, Intercom, Slack, HubSpot, Salesforce or a webhook, and Peeve reads plan and billing state from Stripe and can take a payment action like a refund with confirmation. It also books meetings via Cal.com and Calendly.

### Can Peeve connect to my own API?

Yes, on the Scale plan. Paste your API docs and Peeve builds the connector itself, marking each endpoint read or write. Reads run live to answer; writes run through a confirmation step and an allow-list. The model never sees your keys.

### Does Peeve support MCP?

Yes, both directions. Peeve connects to a Custom MCP server you run on Growth and up, and exposes your capabilities to your users' own AI agents over a scoped MCP endpoint.

### Can AI customer support connect to my own API or backend?

The capable ones can act against your systems, not just SaaS apps. Peeve builds a connector from your API docs on the Scale plan: you paste the docs URL and it maps the base URL, auth and every endpoint, marking each read or write. Reads run live to answer; writes run through a confirmation step and a workspace allow-list.

### Does AI customer support support MCP (Model Context Protocol)?

Yes, MCP is becoming the standard way agents access tools and data. Peeve uses it both directions: it connects to a Custom MCP server you run (Growth and up) to call your tools, and it exposes your capabilities to your users' own AI agents over a scoped MCP endpoint. Anything that changes state runs through a confirmation guardrail.

### Can AI customer support issue a Stripe refund?

Yes, if it can take write-actions on your payment system with a safeguard. Peeve does this on Growth and above through a connector write-action: reads like checking a plan run automatically, while a write like a refund states its effect and runs only after approval, within an allow-list. The model never sees your Stripe key.

### Does AI customer support work with Zendesk and Intercom?

Yes, good agents layer onto the helpdesk you already run rather than replacing it. Peeve resolves in-product questions and hands everything else to Zendesk, Intercom, Slack, HubSpot, Salesforce or a webhook, carrying the full session so the ticket arrives with the investigation done. Most teams run both, and helpdesk volume goes down.

### Can AI customer support read and write to my CRM (Salesforce, HubSpot)?

Reading a contact is common; writing back is what makes it useful. Peeve reads a contact's plan and lifecycle from Salesforce and HubSpot and can push leads and updates, with write operations confirmed before they run. It also reads plan and billing state from Stripe to answer account questions with live data.

### Can one AI agent orchestrate actions across multiple tools in a single workflow?

Yes, resolution often means touching more than one system. Peeve can combine a live read from one connector with a confirmed write to another in the same conversation, for example checking an order and then updating a subscription, with every write confirmed and written to a signed audit log.

### Can AI support read live data, or only static help articles?

Static-only tools answer from a frozen knowledge base and go out of date. Peeve answers from your live product and, when connected, from your real systems: it calls read endpoints during the conversation so the answer reflects the customer's actual account, not a cached article.