This is the practical companion to MCP vs API: what to install, what the AI can do once it is connected, and which permissions to withhold.
The whole setup is three steps, and the third one is the one people skip.
1. Generate a scoped key
In your workspace settings, create an API key for the MCP connection specifically. Do not reuse a key that already runs a Zapier or webhook integration: you want to be able to revoke this one without breaking anything else.
Grant read scopes only on the first key. Leads and companies, campaigns and sequences, inbox threads, pipeline. Leave enrichment, enrolment and sending unchecked. You can issue a second key later, and having two keys with different powers is more useful than one key you are nervous about.
Agencies on whitelabel workspaces should note that keys are scoped per workspace, so connecting one client’s workspace to an AI client exposes nothing from the others.
2. Add the server to your AI client
Every MCP client takes the same two inputs: the server endpoint and your key. Where you put them differs.
- Claude (desktop or web, depending on your plan): add Outboundry under connectors or custom MCP servers, paste the endpoint and key, and it appears as a tool group in the conversation. Claude will list the available tools when asked, which is the fastest way to confirm the connection.
- ChatGPT: add it as a connector in settings. Tool availability depends on your plan tier, so check that the connector is enabled in the specific conversation mode you use.
- Cursor: add the server to your MCP configuration file alongside your other servers. This is the one developers usually want, because it puts outbound data next to the codebase during integration work.
- Anything else: if the client implements MCP, the same endpoint and key work. There is nothing Outboundry-specific to install.
Confirm it is live by asking for something read-only and specific: “list my active sequences with their reply rates”. If the tools are connected you get numbers. If not, you get a polite guess, which is exactly the failure to watch for.
3. Set the limits before the first real task
The order matters. Configure this while the connection is boring, not after an agent has done something surprising.
- Credit cap per session. Enrichment is where money leaves. A mobile number is 10 credits; an agent checking coverage across 500 leads without a cap is a five-thousand-credit afternoon.
- Confirmation on anything irreversible. Sending, enrolment, deal-stage changes on won deals. A confirmation step costs you two seconds and removes the entire category of expensive mistakes.
- No access to suppression or DNC. These are compliance boundaries. Keep them outside the agent’s reach entirely rather than trusting it to respect them.
- Read the tool-call log for a week. Every call is recorded with parameters. This is the advantage MCP has over an opaque assistant, and it is worth actually using.
What to ask it once it works
Prompts that use the connection properly, in rough order of how much trust they need:
| Task | Ask | Needs |
|---|---|---|
| Diagnosis | “Reply rate on the fintech sequence dropped last week. Check volume, bounces and per-mailbox performance and tell me what changed.” | Read only |
| Account brief | “I have a call with this account in an hour. Summarise every touch across email, LinkedIn, WhatsApp and calls, plus the pipeline stage.” | Read only |
| List building | “Find heads of operations at UK logistics firms over 200 people, exclude anyone already in the pipeline or suppressed, and show me the count before revealing contacts.” | Read, then capped enrichment |
| Reply drafting | “Draft a reply to this thread using the last call outcome and the deal stage. Do not send.” | Read + draft |
| Pipeline hygiene | “List deals with no activity in 21 days and propose next steps.” | Read, write on approval |
The pattern in all of them: the agent does the gathering and the reasoning, and a human keeps the irreversible click. That division is not a limitation of the current generation of models. It is how you would structure it for a new hire too.
If something does not work
- The client cannot see any tools. Almost always the key’s scopes rather than the endpoint. A key with no granted scopes connects successfully and exposes nothing.
- It answers without calling a tool. Ask explicitly for live data. Models will happily answer from memory when a tool call is optional.
- Enrichment fails silently. Check the credit balance and any cap you set. A blocked spend looks like an empty result from the model’s side.
- Numbers disagree with the dashboard. Check the date range in the tool call. This is the most common cause and it is almost never a bug.
Next: what agentic outbound actually means, or MCP vs API and who decides the sequence.
Frequently asked questions
What do I need to connect Outboundry over MCP?
An Outboundry account with API access, a workspace API key with the scopes you want to grant, and an MCP-capable client: Claude desktop, ChatGPT with connectors, Cursor, or any other client that supports the protocol.
Which scopes should I grant first?
Start with read scopes for leads, campaigns, inbox and pipeline. Add enrichment with a credit cap once you have watched a few sessions, and add sending last, if at all.
Does the agent spend credits?
Only through enrichment tools, at the standard rates: 1 credit for a verified email, 0.5 for LinkedIn data without an email, 10 for a mobile number. Company records and filters cost nothing, so search and exploration are free.
Can I see what the AI did?
Yes. Every tool call is logged with its parameters and result, and enrichment and sending actions also appear in the normal workspace activity log.
Does this work with a whitelabel workspace?
Yes. Keys are scoped per workspace, so an agency can connect one client workspace without exposing any other.