Connect an AI assistant (MCP)

Let Claude, Cursor, and other MCP-capable assistants work with your Fetchply agents by signing in or using an API key.

Fetchply is an MCP server. Connect an AI assistant once and it can list your agents, check how they perform, surface the questions they could not answer, search what they know, read conversations and captured leads, monitor your plan usage, and even talk to your agents directly, all by asking in plain language. With write permissions it can also maintain your agents' knowledge: add, edit, and remove training sources, refresh training, and manage Q&A pairs.

Tools with a read: permission can never change or delete your data. Write tools act only when the connection has the matching permission: write:chat for ask_agent (uses a message credit and saves the conversation to your Inbox), write:sources for training sources and retraining, and write:qa for Q&A pairs. When you connect with an API key, leave a write permission unchecked to keep the assistant from using it.

Two ways to connect

Sign in with your Fetchply account. For assistants that connect through a sign-in flow, like Claude custom connectors. Add the server URL, and the assistant sends you to the Fetchply login page to approve the connection. The assistant then sees all your agents with every tool available.
Use an API key. For assistants that let you set request headers, like Claude Code and Cursor. The assistant only sees the tools your key has permissions for, and you can limit the key to specific agents.

Either way, the server URL is the same: https://fetchply.com/api/v1/mcp

For the API key option, create the key on the API keys page in your dashboard and select the permissions you want the assistant to have. See creating an API key.

Claude (custom connectors)

In Claude, open Settings → Connectors and choose Add custom connector.
Enter https://fetchply.com/api/v1/mcp as the server URL and save. No key or header is needed.
Click Connect. A Fetchply sign-in window opens; log in with your Fetchply account (or approve directly if you are already signed in).
Once connected, ask Claude about your agents to confirm it works.

Claude Code

claude mcp add --transport http fetchply https://fetchply.com/api/v1/mcp \
  --header "Authorization: Bearer fp_live_YOUR_KEY"

Cursor

Add this to .cursor/mcp.json in your project (or the global Cursor settings):

{
  "mcpServers": {
    "fetchply": {
      "url": "https://fetchply.com/api/v1/mcp",
      "headers": { "Authorization": "Bearer fp_live_YOUR_KEY" }
    }
  }
}

Try it

Ask your assistant: "Which Fetchply agents do I have?" It should call the list_agents tool and show your agents.

Ask: "How is my agent doing this month?" The assistant combines the overview and analytics tools into a summary.

Ask: "What questions couldn't my agent answer this week?" and let it suggest what to train next.

Available tools

Tool	Permission	What it does
`list_agents`	`read:agents`	List your agents with ID, name, and status
`get_agent`	`read:agents`	Read one agent
`get_agent_overview`	`read:analytics`	Performance, demand, and attention items for a period
`get_agent_analytics`	`read:analytics`	Top questions, unanswered queries, satisfaction, locations
`get_unanswered_questions`	`read:analytics`	Questions the agent could not answer
`list_conversations`	`read:conversations`	Page through an agent's conversations
`get_conversation`	`read:conversations`	One conversation with its transcript
`list_sources`	`read:sources`	List training sources with their indexing status
`get_source`	`read:sources`	Read one source, including a text snippet's stored content
`get_training_status`	`read:sources`	Live training progress for an agent
`search_knowledge`	`read:knowledge`	Search what the agent knows and get the most relevant passages
`list_qa_pairs`	`read:qa`	Page through an agent's curated Q&A pairs
`list_leads`	`read:leads`	Page through leads the agent captured
`get_lead`	`read:leads`	Read one captured lead
`list_cod_confirmations`	`read:cod`	Page through cash-on-delivery order confirmations
`get_cod_confirmation`	`read:cod`	Read one cash-on-delivery confirmation
`get_cod_metrics`	`read:cod`	Cash-on-delivery results for a date range
`get_usage`	`read:usage`	Message credits used and when they reset
`add_source`	`write:sources`	Add a URL or text snippet to the agent's knowledge
`update_source`	`write:sources`	Edit a text snippet's title or content in place
`delete_source`	`write:sources`	Delete a source and everything indexed from it
`retrain_agent`	`write:sources`	Refresh the agent's knowledge from its website
`create_qa_pair`	`write:qa`	Add a curated question and answer
`update_qa_pair`	`write:qa`	Change or disable a Q&A pair
`delete_qa_pair`	`write:qa`	Delete a Q&A pair
`ask_agent`	`write:chat`	Send a message to an agent and get its answer (uses a message credit)

For non-Shopify agents on the Free plan, retrain_agent processes up to 100 website pages in each run. Paid plans and Shopify-connected agents do not have this plan-level page cap; their normal training-storage allowances still apply.

Both connection types give the assistant access to your data, including conversation transcripts and lead contact details, and sign-in connections include the write tools. Only connect assistants you trust. To cut off access, revoke the API key from the dashboard, or disconnect the connector in the assistant's settings.

When you connect by signing in, the assistant has every tool. When you connect with an API key, the assistant only sees the tools matching the key's permissions; leave Send messages unchecked if the assistant should not talk to your agents or use message credits, and leave the source and Q&A write permissions unchecked if it should not change what your agents know.

Troubleshooting

The assistant says no tools are available. The key has no permissions selected, or was revoked. Create a key with the permissions you need and update the connection.
A tool reports a missing permission. The key works but lacks that tool's permission. Create a key that includes it and update the header.
The connection fails with an API key. Check the URL is exactly https://fetchply.com/api/v1/mcp, the header name is Authorization, and the value is Bearer followed by a space and your full key.
The sign-in window closes but the connector shows as not connected. Remove the connector and add it again; if you were signed in to a different Fetchply account, log out of Fetchply first so you can sign in with the right one.
The connection stops working after a while. Sign-in based connections expire periodically and the assistant normally renews them on its own. If it does not, disconnect and reconnect the connector.
A tool says "not found or not accessible". The ID it used is wrong, the item was deleted, or the key is restricted to other agents. Asking the assistant to list agents or conversations first usually fixes it.
add_source or retrain_agent reports training already in progress. Only one training job runs per agent at a time. Ask for the training status and try again when it finishes.
update_source says only text sources can be updated. URL and file sources cannot be edited in place. Ask the assistant to retrain the agent for URL sources, or upload the new document from the dashboard for files.
search_knowledge returns an empty list. Nothing relevant is indexed for that topic yet. Check get_training_status; when canChat is true, indexed pages can already answer while the remaining pages continue training. Otherwise wait for the first page, or add a source or Q&A pair covering the topic.
Requests suddenly fail after working. The connection may have hit its per-minute request allowance; the assistant can simply retry shortly.