MCP code intelligence

Agent Memory Products Remember Your Conversation. Nobody Is Remembering Your Codebase

Every memory tool shipped this year stores what you said. Measured sessions of Claude Code auto compaction reduced 132,000 tokens of accumulated state to roughly 2,300, a 98% reduction, and what survives is the transcript rather than the understanding. Multiply that by 40 engineers separately paying a model to work out what the same billing service does, and the cost of not writing the answer down becomes visible.

Waiting for a Bigger Context Window Is a Bet Against Four Orders of Magnitude

At a conservative 1,000 tokens per source file, 10 million files is roughly 10 billion tokens. A 1 million token window holds 0.01% of that, and a 10 million token window holds 0.1%. Closing that by scaling needs something like a 10,000x increase rather than a 10x one, and quality per token moves against you while you try.

Your CLAUDE.md Was Measured, and It Was Roughly Net Negative

An ETH Zurich evaluation found machine generated context files lowered success rates and raised costs by roughly 20%, human written files bought about 4% on the harder benchmark, and Claude Code was the only agent tested where even developer written files failed to beat having no file at all. The most revealing result came from a control: strip the existing READMEs out and context files finally start helping.

The Diff Is the Wrong Unit of Review, and CodeRabbit Is Built on It

A pull request changes 11 lines, the tests pass, every automated reviewer approves, and three weeks later a reporting service is quietly returning wrong revenue. Nothing about the diff was incorrect. What was wrong lived in another repository nobody in the pipeline had read. This is where diff scoped review has nothing to say, and it is now the majority of what breaks in a large organisation.

GitHub Spec Kit Taught Millions of Developers the Right Idea and Left the Hard Half Undone

Spec Kit spent a year and gathered well over 100,000 stars teaching several million developers that the specification, not the code, should be the source of truth. That is category creation at a scale no startup could have afforded, and it lands directly on our thesis. It also assumes a blank page, which is true exactly once in the life of a system.

Repomix Compresses Your Repository by 70%, Which Does Not Solve the Problem

Repomix is the leader in context packing, with 26,000 stars and 255,000 npm downloads a month, and its tree sitter compression genuinely cuts token count by about 70%. Apply that to 10 million files and you go from three orders of magnitude short of the context window to slightly less than three orders of magnitude short. Every tool in this category publishes a token reduction number. None of them publishes a correctness number.

Sourcegraph Answers Where the Code Is. Agents Need to Know Why It Exists

Sourcegraph built the most complete cross repository code search anyone has shipped, and it did so on an assumption that held for fifteen years: a human is asking. Humans supply the meaning the index leaves out. An agent cannot, and at 10 million files a 1 million token window holds 0.01% of the estate, so the gap is not one model generation away.

Context Rot Is Not a Session Hygiene Problem. It Is an Architecture Problem

Compliance with an agent's own explicit constraints falls from 73% at turn 5 to 33% at turn 16 across 4,416 measured trials, and Claude Code auto compaction reduces 132,000 tokens of session state to roughly 2,300. Shorter sessions and cleaner instruction files treat the symptom. The window fills with low value tokens because the agent rebuilds an understanding of your codebase from raw source on every task.

Round-Trip Is a Signal, Not a Promise: How a Verifiable IR Checks Code Against Intent

A technical note on what 'verifiable' actually means in a verifiable code IR. Not a machine that rebuilds your code from a document, but a derived contract that every change gets checked against. Here is what a single verification check consists of, why round-trip is a signal and never a promise, and the honest boundary of what it can and cannot catch.

A Code Graph Disintegrates Your Code. A Verifiable IR Preserves It.

A code graph takes a whole, coherent program and shreds it into a bag of symbols and edges. It is genuinely useful, but it is a teardown, and the meaning never survives the pieces. A verifiable intermediate representation is the opposite move. It lifts your code into a higher layer that keeps the intent intact, then checks every change against it. Here is the difference between shredding and preserving, and why it decides everything downstream.

How to Handle GitHub Copilot's Token Costs Without Going Broke

GitHub Copilot moved to per-token billing and one developer's bill jumped from about $29 to a projected $750. Switching tools won't save you, because the same shift is happening everywhere. Here's why your bill exploded in plain English, the practical levers that cut it today, and how verifiable specs attack the cost at its root.

Your AI Is Charging You Rent to Re-Read the Same Code Every Day

Per-token billing turned a flat $29 AI coding plan into a $750 surprise. The waste was always there: most of the bill is your AI re-reading code it already saw. Here's why it happens and how to cut it, with prompt caching, the right model, the right plan, and a lasting memory of your codebase.

Why Vector Search and GraphRAG Both End in the Same Context Rot

Vector search and GraphRAG look like opposites, but they share the same final move. Parse the code, retrieve some of it, rerank it, and stuff it into the context window. That last step is where both of them rot, because a model that is handed a pile of chunks degrades the same way no matter how the pile was chosen. Here is why the retrieval layer is not the cure for context rot, and what is.

How We Hit 94% Accuracy at 20% of the Tokens Across 46 Repos

The claim that a verifiable IR beats vector search and GraphRAG only matters if the numbers hold up. So here is the run. 46 Kubernetes-ecosystem repositories, 150,000 files, 8 GB of code, measured end to end. What we tested, how we scored it, and why feeding the model 22x more context made it cheaper and more accurate at the same time.

What Verifiable Means for Code Context, and Why GraphRAG Can't Check Code Against Intent

Verifiable is the one word competitors cannot claim. A code graph extracts a shadow of your code and can never tell you whether the code still does what it is supposed to. A verifiable intermediate representation is a derived contract that every change gets checked against. Here is what verifiable actually means, why GraphRAG and vector search can only retrieve, and what continuous verification unlocks.

One Context Layer, Every AI Coding Tool: Serving a Verifiable IR Over MCP

Your team uses Claude Code, Cursor, Copilot, and Windsurf, and each one rebuilds an understanding of your codebase from scratch and shares nothing with the others. MCP is the standard that lets you stop doing that. Compile your repos once into a verifiable IR and serve it to every tool over one MCP endpoint. Here is why the context layer belongs outside the tool, and what changes when it does.

$13.30 to Compile 1,000 Files Into a Verifiable IR, Once

The objection to compiling a codebase into a verifiable IR is always cost. It sounds like Opus pricing across every file. It isn't. At open-source-model rates it runs about $13.30 per 1,000 files, you pay it once, and per-file diffing means you only ever re-pay for what changed. Here is the real economics of the verifiable context layer, and why the expensive thing is not building the IR but living without one.

AST vs Vector vs Graph vs IR: Four Ways to Give an LLM Your Codebase

Every tool that feeds your codebase to an AI coding agent picks one of four representations. AST parsing, vector embeddings, code graphs, or a verifiable intermediate representation. Here is what each one keeps, what each one throws away, and why the first three are features of the fourth.

Spec-Driven Development for Brownfield: Verify Code Against Intent With a Verifiable Code Context Layer

Specs and code drift apart the day after you write the PRD, because they are two separate documents that nobody keeps in sync. A verifiable context layer derives the spec a codebase actually implements and then continuously checks the code against it. Here is how spec-driven development works on a brownfield codebase, and why it needs a verifiable code IR that checks code against intent.

The Context Graph You Can Verify Code Against

Context graph is a term people understand and reach for. But most things called a context graph are one-way extractions you can read and never check anything against. The version worth wanting is the one you can verify code against, and that single requirement is what separates an index from a verifiable code IR. Here is what verification demands, and why it is the only honest test of a context graph.

Code Graph RAG, Explained, and Why a Verifiable IR Is the Next Step Past It

Code graph RAG fixed the biggest flaw in vector search. It follows real edges instead of guessing at similarity. But a graph of symbols is still not a graph of meaning, and it cannot tell you whether a change still does what the system is supposed to do. Here is how code graph RAG actually works, where it stops, and why a verifiable intermediate representation is the layer above it.

Semantic Code Search Is a Retrieval Problem. Context Is a Representation Problem.

Semantic code search keeps getting better at finding the right code, and it keeps disappointing teams who expected better answers from their AI tools. The reason is that those are two different problems. Finding code is retrieval. Understanding code is representation. Better retrieval over a poor representation has a ceiling, and here is why a verifiable code IR raises it.

GraphRAG for Codebases: What It Solves, Where It Breaks, and the Layer Above It

GraphRAG for codebases is a genuine step up from vector search. It follows real call edges instead of guessing at similarity, and it wins on architectural questions. But it breaks in three predictable places: cross-repo links that aren't call edges, the why behind the code, and anything that needs intent you can check changes against. Here is a clear-eyed map of what graph RAG for codebases solves, where it stops, and the verifiable context layer that sits above it.

From Code Graph to Context Graph: Why a Graph of Symbols Isn't a Graph of Meaning

Everyone building code intelligence ends up at a graph, and then they start calling it a context graph because plain code graph stopped feeling like enough. But adding the word context does not add meaning. A graph of symbols and a graph of meaning are different objects. Here is the line between them, and why crossing it requires a verifiable code IR, not more edges.

Why Vector Search, AST Parsers, and Raw LLMs All Fail at Code Intelligence — And What Actually Works

Vector embeddings treat code like english prose. AST parsers see structure but not meaning. Raw LLMs forget everything every session. Here is why the LLM compiler pattern with a persistent semantic graph is the only approach that actually works for cross-repository code intelligence, and why open source models at $7 per 1000 files make it practical today.