Everything this app does, it does through a graph an agent can drive. Claude, Cursor or a script of your own plans lots, prices renders, reads briefs and ships deliveries through the same doors the web app uses.
It renders on your keys, inside your spend cap, and everything it makes carries an agent mark in the editor, so a take you did not personally click is never a mystery later.
Claude, in three steps
LOTGEN speaks MCP at https://mcp.lotgen.ai/mcp. The path matters: the bare origin answers 404, and that code span is a paste target rather than a description of the service. In Claude.ai:
- Add a custom connector at this URL.
- Choose Authentication: always required.
- Sign in and allow it on the screen that opens.
There is no key to copy and nothing to paste into a config file. You sign in as yourself on a consent screen, which is also what makes the next section true.
Sign in and the MCP page inside your workspace has the URL in a copy field.
What a grant actually is
Per client, per person, per workspace.
Allowing Claude does not hand an application a key to your account. It creates one grant, for that client, acting as you, in one workspace. Connecting a second client is a second grant. Joining a second workspace does not carry the first grant over. With the Partner add-on the consent screen can also pick one brand, and the grant then makes and sees that brand's work only, like a key bound to it.
The consequence worth understanding is that an agent has at most your permissions and never more. The consent screen names what the client asked for, in the same words the API keys tab uses (read only, build, build and render, full access), and the grant is held to those. It cannot widen its own access, change a role, remove a member or reach a workspace you are not in — because it is not a separate identity with its own rights, it is you wearing a tag that records which door the request came through.
That tag is what Usage reads to tell you how much of your spend arrived through an agent rather than the web app. An agent never gets a row of its own there, because that would take the spend off the person accountable for it.
Cursor, scripts and anything that cannot sign in
A browser sign-in needs a browser. For a CLI, a cron job or your own code there is a connection key instead, minted from the MCP page once you have a workspace.
It is the same graph and the same permissions; only the way the caller proves who it is differs. Revoke one from that page and it stops working that minute.
Taking it back
The MCP page lists every live grant and every connection key, and disconnects either. Revoking is immediate and it is not reversible: a client you disconnect has to be allowed again from scratch.
Nothing an agent already made is undone by revoking its access. The lots, the takes and the record of what they cost are yours and stay where they are.
Skills, so an agent knows the order to do things in
The tools below are a surface, not a procedure. Knowing that estimate_render exists does not tell an agent to quote before rendering, to pass the references so a silently dropped face gets reported, or that creating a lot spends nothing.
We publish a library of skills that carry exactly that — what to call, in what order, what refuses and why, and which two calls are the only ones that spend. Install them from github.com/ascendedent/lotgen-skills, as a Claude Code plugin or as plain folders any agent can read.
What an agent can reach
Read from the running server, so this cannot list a tool that was removed or miss one that was added.