How John Brahy Turns Claude, Everflow's MCP, and AI Agents Into a Real-Time Network Watchtower
.png)
Most campaign problems don't show up in a spreadsheet until the damage is already done.
By the time someone stitches three reports together, exports a CSV, and builds a pivot table, a payout has already been cut, a tracking pixel has gone quiet, or a scaling affiliate has moved the budget elsewhere. Reporting lag taxes every delayed decision.
John Brahy got tired of asking his team to babysit dashboards, so he wired Everflow's Model Context Protocol (MCP) server into Claude and a set of autonomous agent teams. The result monitors network health, ranks problems by dollar risk, and surfaces a recommended fix, while every action that could actually change something still needs a human to sign off.
Here's how it's built, where it still needs tuning, and how other operators can run the same playbook.
The Architecture: Evidence, Intelligence, and Control
Everflow's MCP server is read-only by design. Models can query live network metrics, attribution rules, and tracking-domain health, but there's no path to modify a record. That's what makes it safe to point autonomous agents at a live revenue system in the first place.
Brahy keeps knowing and doing on separate rails:
Model execution stays flexible on purpose. Heavy reasoning runs on Claude. GopherMind runs the same MCP queries against self-hosted models whenever the data shouldn't leave internal infrastructure at all.
Ditching the Dashboard: What an "AI Harness" Actually Replaces
Standard network management runs on dashboard fatigue. Answering a routine question means configuring a multi-dimensional report and cross-referencing tabs.
"We'd have to merge a bunch of data together and do a lot of those reports with all the dimensions in the analytics section to get the data that we wanted," Brahy says.
An AI harness sits between Claude and both the Everflow MCP and internal database logs. It picks the right tool for the question, pulls live metrics, and returns a merged answer in chat. No dashboard session required.
"It's just part of my environment," Brahy says. "My AI harness grabs the tools when it needs them. It does it for me."
On-demand answers save time day to day. The bigger unlock shows up once the same querying runs on a schedule, unattended.
The Always-On Watchtower, and Where It Still Needs a Human
OpenClaw agent teams continuously compare active offer performance against historical baselines across four areas:
- Protecting earned revenue. Flagging unauthorized payout cuts, tracking cap pacing, catching expired offers still taking traffic.
- Catching silent tracking failures. Clicks continuing while conversions go quiet, isolated through error codes and domain health checks.
- Guarding quality and margin. Cross-referencing revenue, payout, and margin to flag high-volume traffic that's actually unprofitable.
- Surfacing opportunity. Scanning new Marketplace offers for payout fit and automating month-end CRM and billing reconciliation.
None of this works out of the box. Rolling baselines have to account for real seasonality: a holiday traffic spike or a planned promo push looks identical to an anomaly if the model is only compared against a flat 7-day average.
Brahy's agents weight baselines against known calendar events and wait for a sustained deviation, not a single-interval blip, before something earns a spot in the digest. Skip that tuning step and dollar-ranking just reorders the noise instead of removing it.
The Executive Risk Digest
Findings consolidate into one daily digest, ordered by total dollars at risk. Confirmed losses sit at the top, followed by estimated exposure and technical anomalies. Each line item comes with a severity score, dollar impact estimate, supporting evidence, and a recommended fix. It's a decision packet, not a flag.
Team Distribution and Execution Guardrails
Custom API scripts work fine for a technical founder. Scaling this across a team means provisioning access that doesn't depend on everyone knowing how to write a query. Registering the Everflow MCP centrally gives the whole team the same tool access through Claude, with usage logged per person instead of routed through a shared credential.
Non-technical staff use pre-packaged, multi-step agent skills that run a reporting sequence from a plain-language request. An account manager can pull sub-ID performance or verify a payout history without writing code.
Safety Rails
- Least privilege by default. The MCP stays read-only and scoped to the endpoints agents actually need.
- Human-in-the-loop authorization. Pausing offers, adjusting caps, or changing routing requires explicit approval through the separate write path.
- Decision packets, not raw alerts. Every item includes severity, dollar exposure, evidence, and a proposed action.
- Permanent audit trail. Queries and actions are logged centrally, tied to whoever or whatever triggered them.
Compressing the Distance to Decision
The gap between something happening on the network and a human acting on it is where revenue quietly leaks. This setup doesn't close that gap by removing the human. It closes it by making sure the human only has to look at the handful of things that actually matter, with the evidence already assembled. The infrastructure to do this is available off the shelf. The differentiator is in the tuning and the guardrails, not the wiring.
