Case study · Claude · Agency Brain

How we built an Agency Brain on Claude for Contractor Growth Network

Contractor Growth Network knew exactly what its content needed: the real facts about every client. We turned their scattered client knowledge into a per-client brain that the whole team can query from inside Claude, with rules that favour the client's own words and a QA sweep that checks the brain for mistakes every 48 hours.

Client

Contractor Growth Network (CGN), a marketing agency for design-build remodelers

Solution

CGN Brain: a per-client knowledge base served to Claude through an MCP server

Built with

Claude, Model Context Protocol (MCP), SuperMemory, Supabase, Railway, Google Drive, Fireflies

Engagement

E2M Fractional AI Services, from AI Assessment (February 2026) to a dedicated AI developer

01 · Problem statement

Client facts were everywhere except where the team needed them

Contractor Growth Network builds websites, SEO and content for design-build remodelers. Their promise to clients is marketing that reads like a referral, and that only works when every page and every blog gets the details right: the service areas a remodeler wants to own, the projects they are proud of, how they price, what they call things.

Those details lived in Google Drive folders, call recordings in Fireflies, onboarding documents and the memories of the people who ran the calls. When a writer or strategist needed one fact, they had to go digging. When something came out wrong or generic, the client noticed.

CGN had already tried to solve this themselves. Before we got involved, they had built their own retrieval prototype that indexed 48 clients and more than 44,000 chunks of content. It proved the idea, but it was not something the whole team could rely on every day. Ingestion alone had run to a $507 cost overrun, and it needed hardening before it could hold client data safely.

What CGN needed

One source of truth per client that anyone on the team could search in plain language.
Accuracy first: the client's own documents should win, newer facts should beat older ones, and contradictions should be surfaced instead of quietly resolved.
No new app to learn. The team already worked in Claude every day. In March, CGN told us directly that they wanted to use the knowledge from Claude Code and Claude Desktop, so an MCP connection was a big priority.
Security and predictable cost for a system holding every client's data.

02 · The architecture we suggested

Put the knowledge behind Claude, not beside it

Our first proof of concept in March was a standalone question-and-answer app, and it answered client-specific questions such as pricing correctly. But CGN's team did not want another tab to open. So in April we aligned on a different core deliverable: an MCP server.

Model Context Protocol (MCP) is the standard way to give Claude access to outside tools and data. With the brain behind an MCP server, anyone at CGN can ask Claude a question about a client and get an answer grounded in that client's own files, in the same Claude workspace where they write, plan and review.

01

Claude is the interface

No dashboard to adopt. The brain appears as tools inside Claude Desktop, Claude Code and Claude Cowork.

02

One memory container per client

Each client's knowledge is kept in its own container, so answers stay specific to that client.

03

Rules before retrieval

Source priority, freshness and conflict handling are designed in, not left to the model's judgement.

Sources

Client folders on Google Drive shared drives
Call transcripts from Fireflies
Client feedback and corrections

Ingestion

Sync workers on Railway
Transcript2Drive workflow files every call to the right client
Drafts, archived clients and sales prospects excluded

Knowledge

SuperMemory: one container per client
Supabase: client mappings, name history, conflicts
Source ranking and time decay

Claude

CGN Brain MCP server, 36 tools at launch
Sign-in restricted to the CGN domain
Claude Projects per client and skills such as /rag-query

Quality loop: a QA Sweep plugin running in CGN's Claude Team plan asks the brain questions with known answers every 48 hours and files anything that fails to a shared Monday.com board.

Why this stack

Secure by design

An early security review in April found that the original prototype could be manipulated into revealing API keys. That finding drove the move to a managed memory layer behind an authenticated MCP server.

Small running cost

At launch the core infrastructure was SuperMemory at about $20 a month plus a $20 Railway plan, about $40 a month in total.

Fits how CGN works

Google Drive stays the place files live, Fireflies keeps recording calls, Monday.com tracks issues and Claude is where the work happens. Nothing had to be replaced.

03 · The solution we implemented

A brain the whole team can ask from inside Claude

01

Hardened the brain and connected it to Claude (April)

We rebuilt the knowledge layer on SuperMemory and Railway and exposed it as the CGN Brain MCP server, with 36 tools for listing clients, querying knowledge, handling conflicts and more. By the end of April, anyone with a Contractor Growth Network Google account could sign in, Google Drive changes synced into the brain automatically (today this runs every hour), conflicts were logged, and 50 clients were ingested.

02

Taught the brain which facts to trust (April and May)

Accuracy became the top priority, and CGN approved ingesting every client in full. We then added the rules that make answers dependable:

Client sources first

Client documents outrank Google reviews.

Newer beats older

Time decay went live in May, so outdated facts lose weight.

Conflicts flagged

When sources disagree, the answer says so instead of picking one.

Noise kept out

Files marked draft, archived clients and sales prospects are excluded.

03

Made every call land in the right client's brain (May and June)

We rebuilt the transcript flow as Transcript2Drive V2, using CGN's Fireflies admin access so every user's calls are captured, and added a 120-second wait to stop silent failures. It went live in early June.

04

Turned CGN's own QA idea into a team-wide Claude plugin (May and June)

Logan Shinholser, CGN's President, had built a first QA plugin himself in Claude Cowork. E2M took ownership in May and released it as QA Sweep in CGN's Claude Team plan in June. Every 48 hours it runs canary questions with known answers and checks for problems like duplicate clients, then files findings to Monday.com. CGN's team wrote human-verified canary answers for 10 clients to anchor it.

05

Trained the team to work through Claude (May, then September)

On 29 May we ran a hands-on Claude and MCP session for CGN's team, and 12 CGN team members accepted the invite, from leadership and project management to writing and SEO. The working model we set: a Claude Project for each client, and skills for repeatable tasks. In September we set up Claude Projects for CGN's priority clients, each routing client questions to the brain through the /rag-query skill.

06

Scaled it for heavier use (August and September)

As more of CGN's workflows came to depend on the brain, we moved the MCP server to an async client with a job handoff for long QA requests, and added an admin view of Claude usage against the brain by user.

04 · Notable outcomes

The brain became the foundation CGN builds on

Every figure below comes from our weekly sync calls with Contractor Growth Network, their written updates and our project records.

50

clients ingested into the brain

In the production brain by 29 April 2026

300–400

content briefs generated on brain facts

Reported by CGN on 8 September. Brief facts are built only on brain answers

48 hrs

between automated QA sweeps

QA Sweep live in CGN's Claude Team plan since June

~$40

per month core stack at launch

SuperMemory plus Railway, replacing a prototype that overran $507 on ingestion

01

The team rebuilt its processes around it

Within weeks of launch, CGN's leadership described the brain as the base layer for how the team works.

Our team is now redoing their processes where the brain is like the foundation of it.
Logan Shinholser, President, Contractor Growth Network, May 2026

I always say use the brain. I always, I pull up the brain first.
Logan Shinholser, June 2026

02

It powers the work that ships to clients

The brain is what makes CGN's AI content trustworthy. Content briefs pull their client facts only from brain answers, and CGN reported 300 to 400 briefs generated across their SEO clients by September. Writers draft blogs against it, and CGN's project management team uses it to QA work before it goes out.

03

Accuracy problems get caught by a system, not by clients

The QA sweep has already surfaced issues such as duplicate client records and off-boarded clients still sitting in Drive, so they could be cleaned up before they turned into wrong answers. Fixes land on a shared Monday.com board where they are tracked to done.

04

Claude knows each client out of the box

With a Claude Project per priority client and the /rag-query skill, a CGN team member can open the client's project and ask a question in plain language, and Claude answers from that client's own files.

What we are still improving

A knowledge base is never finished, and we would rather be clear about that. Detailed answers can take 30 seconds or more to generate. Heavy parallel use exposed concurrency limits, which we are addressing. Usage quotas on underlying services caused interruptions, and because CGN's team relies on the brain every day, they felt them. Occasional wrong facts still surface in drafts. When they do, they are traced to the source and corrected in the brain. The next steps are hardening the infrastructure and feeding client feedback from CGN's portal straight back into the brain.

Key takeaways for agencies

What we would tell any agency building a brain on Claude

Meet your team where they already work. An MCP server made the brain part of Claude instead of another tool to adopt.
Accuracy is a design choice. Source ranking, freshness and conflict flags matter more than which vector store you pick.
Test the brain like a product. Known-answer canary questions on a schedule catch problems before clients do.
Security review early. A prototype that holds every client's data deserves a proper review before the whole team relies on it.
Build on what the client already made. CGN's own prototype and QA plugin shaped the final system. Our job was to make them production ready.

Want an Agency Brain on Claude for your team?

E2M's Fractional AI Services start with an AI Assessment and Blueprint, then put a dedicated AI developer on the systems that matter most to your agency.

Talk to E2M about Fractional AI Services