Python 3.10 or newer. The example repository runs these steps on a fresh machine in under a minute.
git clone https://github.com/MoltyCel/moltrust-ads-example cd moltrust-ads-example python -m venv .venv && . .venv/bin/activate pip install -r requirements.txt export MOLTRUST_API_KEY=... # free key: see Pricing python quickstart.py
Expected: one allow, one hold, two forbid, each with a record digest.
| Field | Meaning |
|---|---|
tools | Allowed Amazon Ads MCP tools, by exact name. Everything else is forbidden. |
account_ids | Ad account or profile IDs (accessRequestedAccount). |
brands | Brands the agent may touch. |
daily_budget_cap, hold_threshold | Cents. Above the threshold: hold; above the cap: forbid. |
total_daily_budget_cap | Cents. Sum of daily budgets on the account after the change. |
period_budget_cap | Cents. Lifetime and monthly budgets. |
max_bid, max_bid_change_bp | Cents; basis points per change (2500 = ±25 %). |
valid_from, valid_until | UTC. Outside the window every call is forbidden. |
ratification_authorities | People who may approve a hold, with their public keys. |
The guard compares new values against your view of the account: campaigns (brand and current daily budget per campaign ID), targets and ad_groups (brand, current bid), and account_daily_total. Take these from your own reporting store, not from the model.
For a hold, pass an approver to the guard. It receives the pending record, gets a signature from an approver named in the mandate and returns it; the guard sends it to POST /enforce/ratify and runs the call only on RATIFIED.
moltrust-verify --input decision.json --trust-list trust.json V1 PASS bundle commit and input binding hold V2 PASS ... V4 PASS recomputed DENY from the same inputs
Every check carries an AER record. The verifier needs no network and no MolTrust account. Reference: moltrust-enforce SDK, ads profile encoding.
An agent can onboard itself. On the hosted MCP endpoint https://api.moltrust.ch/mcp: ads_onboard returns plans, sign-up and endpoints; ads_mandate_template builds the profile and mandate from account IDs, brands and caps in USD; ads_check returns allow, hold or forbid for one intended Amazon Ads tool call. ads_check needs your MolTrust key as the X-API-Key header of the MCP connection and refuses anything shaped like an Amazon credential. The same skill is listed in the signed agent card as ads-verdict-gate.