Look What GitLab Invented!

#aitokencosts #gitlab #claudecode #jcodemunch

The industry just got slapped awake by two numbers it cannot unsee.

Uber torched its entire 2026 AI coding budget in four months. Its COO went on record saying he still cannot draw a straight line from all that Claude Code spend to shipping more useful features. "That link is not there yet."

At the same time, an unnamed company accidentally dropped half a billion dollars on Claude in a single month because nobody put a cap on employee licenses.

The panic narrative writes itself:

"AI IS TOO EXPENSIVE!"

The diagnosis everyone is missing: most of those tokens were spent forcing models to skim entire files just to find the three functions, classes, or references they actually needed.

You are not paying for intelligence. You are paying retail, per token, to make the robot read the whole damn book when it only needed one page.

GitLab just validated the fix in public

While the budget fires were burning, GitLab was building the same fix in the open. They call it Orbit, their Knowledge Graph. It turns repositories into a structured, queryable map of definitions and cross-file references, then hands agents only the precise context they ask for through what they literally call a unified context API. It even exposes an MCP endpoint.

Sound familiar?

It should. That is the entire premise behind the jMRI spec and jCodeMunch.

GitLab's own Knowledge Graph team wrote down exactly why they built it. In their internal engineering notes they describe treating AGENTS.md as a table of contents instead of an encyclopedia, keeping knowledge in structured docs, and enforcing architecture mechanically. Their stated design goal is agent legibility, on the logic that anything the agent cannot access in-context effectively does not exist.

A four-person team used that discipline to ship roughly 135,000 lines of Rust in about two weeks, around 95% of it agent-generated. They validated the thesis with their own velocity numbers.

I have been screaming that from the rooftop since March.

Here is where the big-vendor version gets... funny

GitLab's full SDLC graph, the one that actually maps merge requests, pipelines, work items, and code together, is Ultimate tier, GitLab.com only, runs in its own Kubernetes cluster, and is still sitting behind an experiment flag that is off by default. The local code piece writes to a private DuckDB file on your machine and is built to feed GitLab's own stack.

The insight is correct.

The delivery vehicle is "buy our most expensive seat, stand up a cluster, and put your source on our infrastructure."

The same idea, on your laptop, for the price of a tank of gas

jCodeMunch does the retrieval part locally. It is MCP-native, so any agent that speaks MCP can use it. It works against any repository on any host. Your code never leaves your machine. No recurring enterprise seat. No cluster to babysit.

This is not theory. Universal Plant Services, an industrial-services company, is already running the full jMunch platform across more than forty engineer seats. Their live dashboard, as of this week:

Nearly 48.7 billion tokens saved across the fleet in 72 days, an API-rate value of $177,256.07.

That is what jCodeMunch's savings counter put up, in a couple months, at a single company.

While other organizations are writing postmortems about leaderboards that rewarded waste, a real forty-plus-seat shop is quietly running the efficient version and letting the numbers speak.

Honest side-by-side

Aspect	GitLab Orbit	jCodeMunch
Where it runs	Their Kubernetes cluster, GitLab.com only	Your laptop
Price	Ultimate-tier seat, ongoing	One time, per seat or platform
Repositories	GitLab-hosted, default branch only	Any repo, any host
Model lock-in	Built for GitLab's own stack	Model-agnostic over MCP
Your source code	Lives on their infrastructure	Stays local
Status	Experiment, feature-flagged, off by default	Shipping, running at 40+ seats today