Claude Code, pointed at your ad accounts

Claude Code runs in a terminal and holds a project's context, which makes it an odd place to run advertising until you notice that most ad work is the same shape as most engineering work: read the state, decide on a change, write it down, have someone check it, apply it, verify it landed. The gap has always been that ad platforms have no such workflow. This one does.

Tempio's programmatic surface is a Model Context Protocol server over streamable HTTP. There is no separate REST API to learn: the in-app agent and your own client call the same org-scoped tool layer, so a tool behaves identically whichever one is holding it. The endpoint takes a bearer key you mint in the app, and the workspace is read from the verified key rather than from anything a tool is asked for — which is the whole of the multi-tenancy story, and the reason a key cannot be talked into reaching a workspace it was not minted for.

Keys carry scopes. Reading covers connections, ad accounts, campaigns, structure, media assets, metrics, drafts and media plans. Drafting creates change requests, edits them, undoes applied changes by drafting their inverse, and edits plans; nothing it writes reaches a live ad account. Approving applies or rejects a pending draft and is opt-in at the moment the key is minted. Scopes are enforced on the call rather than by hiding it: a read-only key asked to draft is refused by name, with the scope the capability needs and the scopes the key holds, so an agent is told what to ask for instead of concluding the capability does not exist.

Mint the key first, under Settings and then MCP in the app. It is shown once, hashed at rest, revocable, and scoped to one workspace. Then add the server:

claude mcp add --transport http tempio https://app.tempio.ai/api/mcp \
  --header "Authorization: Bearer YOUR_KEY"

Where that configuration is written is a decision worth making deliberately. The default scope keeps it private to you for the current project; user scope makes the server available across all of your projects, still private; project scope writes it into a file in the repository root, which is shared through version control and is the wrong home for a bearer key. For a key, use the default or:

claude mcp add --transport http tempio https://app.tempio.ai/api/mcp --scope user \
  --header "Authorization: Bearer YOUR_KEY"

Then check the handshake rather than assuming it. Listing the configured servers from the command line shows whether this one connected; the same check inside a session shows the server and the tools it exposed. What the tool list itself shows is the surface chosen at the mint: compact registers two tools, one that describes the catalogue and one that runs it, and explicit registers each capability as a tool of its own. Both reach the same capabilities under the same scopes, so the list is a context-cost decision and the refusal is the gate.

Start with the account brief. It is ninety days of context per connected ad account — coverage, totals, a twenty-eight day period against the one before it, the objective mix, how concentrated spend is, which campaigns rank on return and which rank worst on cost per acquisition, and which weekdays are cheapest — and it exists so that an agent has a baseline before it calls a number good or bad. It says when it was computed, and it says so out loud when it is more than a week old.

After that, the reads compose the way you would expect. Blended KPIs with a per-platform breakdown for a date range. Per-campaign rows highest spend first, each in its own account's currency, carrying a marker that says whether the order is a true ranking or several currencies laid side by side. The structure beneath any campaign or ad group, each row carrying the operations that level can actually express on the platform it lives on. The media library, each asset with the id the platform holds those bytes under.

Two rules run through all of it. Money is integer minor units with an explicit currency, so nothing is divided by a hundred on a hunch. Ratios are computed when they are read from base counts, and a ratio with a zero denominator is null rather than zero — a campaign with spend and no conversions has no ROAS, and an agent that reports zero is reporting something that did not happen.

Ask for a change and you get a change request. One call writes a whole batch in the order it applies, each entry able to name the earlier entry that creates its parent, so a campaign, an ad group and an ad arrive as one reviewable group rather than three acts that can half-succeed. What comes back is the batch, one review link for the group, and the authority that cleared it: waiting for a person, or cleared on the way in by a rule the workspace wrote.

The client's job at that point is to hand you the link rather than to keep going. A draft that needs an edit has its payload replaced whole, which re-validates it, re-stamps the currency from the target account and re-runs the guardrails — so a stopped batch is resumed rather than written again from scratch. A change that already landed is undone by drafting its exact inverse at the same gate, which means an undo is a change request like any other rather than something happening behind you.

Approving is deliberately not the same permission as drafting. Most workspaces hand their agent a key that can read and draft, and keep approval for a person or for a rule they wrote. The record says which of those cleared each row, and it never pretends a policy was a person.

The same setup run on one platform, end to end is the walkthrough for one platform, with the reads and the drafts that platform actually accepts.

It cannot see another workspace. It cannot approve anything with a key that was not minted with that permission. It cannot upload a file into an ad account's media library — the library lists what is there, and a creative references an asset by the id the platform already holds it under. And it cannot make a change take effect without something clearing it, which is the point rather than a limitation: a queue of drafts with a review link is a smaller problem than a change nobody remembers making.

The approval gate is where those drafts wait, the media plan is where a flight is shaped before any of them exist, and the MCP server page is the same story from the product's side rather than the client's.