# Derek Sato, Senior Platform Engineer at Fenix Payments — read of Codebase Memory MCP, June 24 2026

> 11 years in backend and infra, last 18 months deep in LLM tooling. We're running agentic code review pipelines on Claude and the API bill is becoming a weekly topic in standup.

## How I got here

Someone dropped a link in the MCP Builders Slack channel I lurk in. The thread was about token cost management for long-running agents. Three people had commented "interesting" without any follow-up, which is either a good sign or the internet's way of being politely unimpressed. I clicked through during lunch.

## What I clicked first

"99% token reduction for AI agents querying your codebase." That number is aggressive enough to make me either interested or suspicious. Then I saw the comparison: "~45,000 tokens per agent turn (large repo)" vs "~450 tokens per agent turn (cached)." I stopped on that. What counts as a large repo? Our monorepo is 380k lines. I have no idea if their benchmark means a 5k line toy app or something real.

## Where I paused

The pricing model: "Pay for repositories indexed, not queries run." That's actually a thoughtful structural choice and it caught my attention. Most tools charge per query and you end up metering yourself into bad behavior. If I could run 10,000 agent queries against an indexed repo for the same $29/month, that's a real model. I paused because it's either a strong conviction about the value prop or it breaks at scale in ways they haven't hit yet.

## What I distrusted

The page starts looking like a live SaaS product hero. "Start Free Trial." Pricing table. Feature bullets. Then near the bottom it says "Honest disclosure: we don't have live customers on this idea yet. We shipped the strategy package; you ship the customer conversations." That is a jarring gear shift. I had to re-read the whole pricing section to figure out whether I was buying a subscription to a real product or buying a $99 dossier to go build this myself. Turns out it's the latter. But the hero section doesn't tell me that. The page is designed like a product and reveals at the bottom that it's an idea for sale. That's a trust issue. Not a dealbreaker for what this actually is, but it made me feel briefly misled.

The line "Agents spend less time reading code, more time solving problems. Your API bills go down. Your agent reliability goes up." reads like someone told an AI to write benefit bullets and didn't edit them.

Also: the page's own scoring system gives the landing page a 3/10 on quality. They published that. Points for honesty, but also: fix the page.

## What would convince me

A real benchmark against a specific, named, public repository with a stated model and agent framework. "We ran a code review agent against the FastAPI GitHub repo (X files, Y lines) and here's the token count before and after, logged from the Anthropic dashboard." That's the thing I would forward to my team. Not a comparison table with round numbers and no methodology footnote.

Also: one paragraph explaining what the MCP resource actually returns. Does it return AST nodes? Summaries? Function signatures with docstrings? That changes everything about whether it's useful for my specific agent pattern.

## What I'd ask in an email reply

1. The 45k to 450 token comparison -- what codebase was that, what model, what agent loop? I need to reproduce that estimate against our stack before I can take it to anyone internally.
2. How does incremental indexing work when we push 30 commits a day to a feature branch? Does it re-index on push, on schedule, or on first query?
3. You're at zero live customers -- are you looking for design partners who get early access in exchange for feedback calls, or is this purely a pay-and-build model?

## Verdict: on-the-fence

The core problem is real and I live with it every week. But I walked away confused about whether this is a product I can trial today or a business kit I'm supposed to go build. If it's the former, show me a working demo repo. If it's the latter, say that in sentence one.

---
*Memo by skeptic persona, generated 2026-06-24. Studio breaks own self-grading loop.*