What a custom MCP server is, and what it is not
MCP (Model Context Protocol) is an open way for a program to publish tools that an AI agent can call. The service runs a server; the agent is a client. Presumption is that client: it opens a short session with your server, asks what tools it offers, and uses them for the job you described.
That is the important limit. A custom MCP server is not a general API runner. Presumption will not POST arbitrary JSON at any endpoint you hand it. The service has to speak MCP - the same initialize-and-list-tools handshake every MCP client uses - so the agent is always choosing from a tool list the server itself published, with the input schema the server itself declared.
The practical upside is breadth. Your own internal service, a niche vendor nobody has built an integration for, or a self-hosted tool can all become part of a workflow, without waiting on anyone's roadmap.
- The server must be reachable over the public internet at an https address.
- Authentication has to be a header the server accepts, such as an API key. Servers that require a browser sign-in (OAuth) are not supported yet.
- You are granting the agent whatever that server allows its key to do, so the key's own scope matters.
Before you paste a URL: three checks
Most failed connections are one of three things, and all three are quick to check first.
- Is it https and public? Presumption refuses plain http, localhost, and private network ranges such as 127.x, 10.x, 192.168.x and 172.16-31.x. A server running only on your laptop needs a public tunnel or a real deployment.
- Can it authenticate with a header? Use a key or token the server accepts in a request header, for example Authorization: Bearer <key> or X-API-Key: <key>. If the only way in is an OAuth browser flow, that server cannot be added today.
- Do you trust the tool list? Reading the published tools tells you what the agent will be able to attempt. A server that exposes delete, refund, or send-message tools is granting those actions, subject to your approval settings.
Step by step: connecting the server
Open Settings, then the Connections tab, and scroll to Custom MCP servers. The panel is honest about the rule that matters: lookups run on their own; anything that changes data waits in Approvals unless the automation has auto-approve on.
Choose + Add an MCP server and fill the four fields. The name is yours to read later, so make it obvious which service it is. The address is the server's MCP endpoint. The header name defaults to Authorization; change it only if your server expects something else. The header value is the secret itself - optional, and only needed when the server requires a key.
Choose Connect server. Presumption opens a session, asks the server for its tools, and only saves the row if the handshake worked. If the server answers but publishes nothing, you get 'Connected, but that server doesn't offer any tools.' - the address was fine, the tool list was empty. If a name is already taken in your account, you'll be told to pick another.
Once it is saved, the row shows the server's name, how many tools it published, and its host. The refresh icon re-reads the tool list after the server changes; the trash icon removes the connection.
- Name: a short label you will recognise, such as 'Acme Support'.
- Address: the https MCP endpoint, for example https://mcp.acme.com/mcp.
- Header name: usually Authorization. Plain names only - host, cookie, and a few other reserved headers are refused.
- Header value: your key or token. It is stored encrypted and is not shown again afterwards.
- Adding a server is available on paid plans, and each account can connect more than one.
Read the tool list before you use it
Choose the tool-count line on a saved server to expand its published tools. Each row shows the tool's name, its description, and a tag: read, or needs approval. That tag is the permission surface you are handing the agent, so read it before pointing any workflow at the server.
The tag comes from the server's own declaration of which tools only look things up. Treat a mismatch as a warning about the server, not a bug to work around: if a tool that clearly writes data is marked read-only, that server is mislabelled and you should not grant it auto-approve.
Re-read the list after the service updates. Refresh re-fetches it, so a server that added a destructive tool since you connected will show up.
Point the agent at it
With a server connected, describe the job the way you would to a colleague. When your account has exactly one server, the agent uses it. With several, name the one you mean - the automation carries the server's name or id, and a run that cannot find it stops with 'Add the "<name>" MCP server in Settings -> Connections to run this.' rather than guessing.
Under the hood, Presumption hands the agent the server's tool catalogue, the tool's input schema, and your task, and the agent picks exactly one tool plus its arguments. That keeps the choice inside what your server actually published instead of an invented call.
The agent brief
Use the MCP server named Acme Support. First list its tools and tell me which are read-only and which need approval, so I can see the permission surface. Then use a read-only tool to pull the open support tickets from the last 7 days and summarise them in 3-5 plain lines: what keeps recurring, which customers are affected, and what looks most urgent. Do not call any tool that changes data in this run.
What runs on its own, and what waits for you
This is the part worth understanding before you schedule anything. A tool the server marks read-only runs immediately, and its result is summarised into the run's output. Anything else does not fire: it becomes an entry in Approvals showing the server, the tool, and the exact arguments the agent chose, and the run pauses at 'needs approval' until you decide.
Auto-approve on the automation card changes that. With it on, the action is carried out without waiting for you - the same rule that governs email and social posts. That is genuinely useful once a workflow has earned it, and it is real permission, not a shortcut for skipping a draft.
- First run with a new server: keep auto-approve off and watch what the agent asks to call.
- Approvals show the arguments, so you can see the record it intends to touch before you agree.
- A read-only lookup that fails is reported as a failure in the run, not silently skipped.
- The same approval policy applies whether the run is on demand or on a schedule.
Good first jobs for a server you own
Start where the work is boring, repeatable, and safe to look at. Read-only jobs prove the connection and show you how the agent reasons about your data before it is allowed to change any of it.
- Lookups: open tickets, order status, inventory levels, account balances, campaign spend.
- Reporting: pull the same numbers every morning and have them summarised into a daily brief.
- Sync: bring records from an internal system into the CRM so follow-up happens somewhere you already work.
- Drafting from live data: write a reply or a renewal note from what the server actually returned, then hold it for your review.
- Writes you supervise: closing a ticket, updating a record, queuing a send - all fine once you have watched the agent choose correctly a few times.
When it does not work
The messages are specific on purpose, because each one has a different fix.
- 'That isn't a valid web address.' - the value is not a URL. Include the scheme, for example https://mcp.acme.com/mcp.
- 'The MCP server address must start with https://' - plain http is refused.
- 'That address points at a private network, which isn't allowed.' - the host is localhost or a private range. Expose the server publicly or use a tunnel.
- 'Use a plain header name like Authorization or X-API-Key.' - the header name contained characters outside letters, numbers, and hyphens.
- 'That header can't be used for authentication.' - reserved headers such as host, cookie, and content-length cannot carry your key.
- 'Connected, but that server doesn't offer any tools.' - the handshake worked and the tool list was empty. Check the server's configuration, then refresh.
- 'You already have a server with that name.' - pick a different label; the address itself is fine.
- 'Couldn't pick a matching tool on <server> for this task.' - the task did not map to any published tool. Name the tool you want, or widen the server's tool list.
- Long waits that end in a failure - the session is given about 25 seconds. A server that answers slowly needs its own performance fixed.
What this replaces, and what it does not
It replaces the glue. The usual alternative to a missing integration is a script nobody owns, a scheduled job that breaks silently, or a person doing the copy-and-paste. A server you connect in thirty seconds sits inside the same system as everything else: it runs on a schedule, it hands its output to the next step in the chain, and every run keeps its result and history.
It does not remove your judgment or your obligations. The key you paste is yours, and so is whatever that key can reach. Truthful claims, lawful targeting, and care with customer data stay with you, exactly as they do with email and social publishing. And a tool list is not a promise: a server that publishes a destructive tool has published it, which is why the approval step exists.
The honest summary is that Presumption stops being a fixed list of integrations. You decide what the agent can reach, you read what it can do, and you decide what runs on its own.

