Blog · methodology, evidence, experiments
Writing from the project.
Long-form notes on ad-ops methodology, security posture, and what we've learned running mureo against real accounts. Evidence-based and data-forward; no hype, no lists of five tips.
-
Bringing ad ops in-house: what comes back, and what newly costs you
Switching ad ops from an agency to in-house removes the management fee, but the hours and the responsibility for judgment move to you. A table comparing fees against internal hours in yen at each level of spend, what happens to the account and the learning phase when you switch, the three things to take with you, and how to choose not to bring everything in-house, from where we sit building ad-ops tooling.
-
Can connecting your ad account to Claude get it banned?
The rumour says accounts get suspended for wiring ads into an AI agent. What the Google and Meta terms actually prohibit, which route puts the account at risk, and the checks to run before you connect a client.
-
Driving Amazon Ads from Claude Code
How Claude Code operates Amazon Ads over MCP and the Amazon Ads API, and the Amazon-specific places that trip people up — profile IDs, asynchronous reports, and delivery stopping because of stock — from where we sit building ad-ops tooling.
-
Driving Google Ads from Claude Code
How Claude Code operates Google Ads, taken in order: MCP, the Google Ads API, the developer token, GAQL. Plus the places people actually get stuck connecting, from where we sit building ad-ops tooling.
-
Outsourcing Google Ads management: costs and Google-specific cautions
Outsourced Google Ads management in the Japanese market runs around 20% of ad spend. But Google Ads has particulars worth knowing before you delegate: automated bidding, Performance Max, account ownership, auto-applied recommendations, and what the fee actually buys — from where we sit building ad-ops tooling.
-
Outsourcing search ad management: costs, and how to pick a provider
Outsourced search ad management in the Japanese market runs around 20% of ad spend on top of the media cost, with pricing splitting into rate-based, flat, and performance-based. What each suits, and what to verify before choosing a provider, from where we sit building ad-ops tooling.
-
Driving Meta Ads (Facebook and Instagram) from Claude Code
How Claude Code operates Meta Ads across Facebook and Instagram, seen through the relationship between MCP and the Meta Marketing API. Plus the places that trip people up, measurement and the learning phase in particular, from where we sit building ad-ops tooling.
-
Driving Yahoo! JAPAN Ads from Claude Code
How Claude Code operates Yahoo! JAPAN Ads (search and display) over MCP, and where the current shortage of tooling compared with Google Ads trips people up, from where we sit building ad-ops tooling.
-
Are ad agency management fees too high?
Management fees in the Japanese market run around 20% of ad spend, and the smaller the budget the more expensive that gets. This post breaks down how the fee is set and how costs change if you in-house, move to a flat rate, or move to AI ad ops, from where we sit building ad-ops tooling.
-
Reading your agency's ad report: the numbers to check first
The monthly report from your agency arrives and the numbers do not say much, so the unease of having delegated everything stays. The metrics an advertiser should check first, what a report tends not to show, and how to make the substance visible while still delegating, from where we sit building ad-ops tooling.
-
Can you automate ad ops with Claude Code?
What actually works when you automate ad ops with Claude Code, and where the pitfalls are: data pulls, reporting, anomaly detection — and the design decisions worth settling before you hand anything over, from where we sit building ad-ops tooling.
-
Is it safe to hand ad ops to Claude Code?
The risks of delegating ad ops to Claude Code — bad operations, hallucination, runaway execution, lost accountability — and the design that keeps them in check, from where we sit building ad-ops tooling.
-
Why ad ops becomes key-person dependent
Running multiple clients at an agency makes quality vary by operator and decay at every handover. This post breaks down why key-person dependency happens, and works through the design — environment isolation, criteria, permissions, records, two-layer knowledge — from where we sit building an ad-ops platform.
-
Insight federation: how mureo borrows practitioner know-how without leaking it
mureo 0.9.19 ships `mureo_consult_advisor`, a federated retrieval surface that lets diagnostic agents pull practitioner know-how from external advisor servers without anyone shipping their corpus out. This post explains the v0.9.18 design that leaked and the v0.9.19 retrieval pattern that does not.
-
mureo as a control plane: the plugin architecture in 0.9.x
mureo 0.9 ships in ten releases that together rebuild the framework as a pluggable control plane: a stable Capability+Protocol ABI, entry-points discovery, an ABI stability promise, configure-UI web extensions, and a safety layer that wraps every third-party plugin tool. This post walks the architecture and explains why this shape matters once the official ad-platform MCPs ship.
-
Try mureo without an ad account: anatomy of `mureo demo init`
mureo 0.8.0 ships four synthetic scenarios that let an evaluator run the agent end-to-end without OAuth, an XLSX bundle, or even a real ad account. This post unpacks what each scenario is designed to surface and why a synthetic dataset is the right shape for evaluating ad-ops AI.
-
Open source ad ops: an Apache 2.0 commitment
Why mureo is open source under Apache 2.0, what stays on the operator's machine, what the project will not do, and what that means for agencies and in-house teams evaluating long-term risk.
-
Anatomy of /daily-check
What mureo does when an operator types a single three-word command: four-platform data pull, strategy binding, anomaly detection, and a single correlated report — with a walkthrough of the manual equivalent.
-
Running mureo as a sidecar container
When the local pip install is not the right fit — CI runs, agency multi-tenancy, non–Claude Code MCP clients — the new Dockerfile gives mureo a sidecar shape without giving up the local-first credential posture.
-
Strategy-driven vs metric-driven optimization
Why metric optimization alone misallocates spend, what lives in STRATEGY.md, and a worked example where two accounts with identical numbers require opposite actions.
-
The threat model of AI agents touching ad accounts
Where ad-ops AI agents can be exploited — prompt injection, credential exfiltration, unbounded mutations — and the four mechanisms mureo uses to contain the blast radius.
-
How AI agents misdiagnose CPA spikes
Why naive threshold alerts on ad cost-per-acquisition fail at the point they are most needed — and the three mechanisms mureo uses to fire only on signal.
New posts land weekly. Subscribe via RSS.