Why your AI coding agent burns tokens — and how a code map fixes it

The hidden cost: agents read the whole repo to answer one question

An AI coding agent — Claude Code, Cursor, or any other — can only reason about code it has loaded into context. To answer “where does authentication happen?” it first has to find the auth code, and without a map of the codebase it does that the only way it can: it lists directories, greps for keywords, and opens files until it has seen enough. A typical 500-line file is 3,000–5,000 tokens, and exploring an unfamiliar repository routinely burns 50,000+ tokens before any real work begins.

It gets worse over a session. Most agents re-send the entire transcript — every file already opened — on every turn. So token usage grows with the size of the codebase and the length of the conversation, not with the difficulty of the task. That is how a single afternoon of agentic coding turns into a surprising bill.

Why a bigger context window isn't the fix

The instinct is to throw more context at the problem. But a larger window only lets the agent hold more files — it still has to discover which files matter, and packing the window with marginally-relevant code measurably degrades answer quality while costing more per turn. The lever that actually works is the opposite one: send less, more precisely. The right neighborhood of code beats the whole repository every time.

A code map gives the agent the right neighborhood, not the whole repo

A code map — also called a code knowledge graph — is a pre-computed index of every file, symbol, import, and call in your project. Instead of crawling, the agent queries it: “what does this task touch?” returns a focused, ranked slice of code with exact path:line anchors. The agent reads only those slices. Discovery stops being a token sink and becomes a lookup.

What it looks like in tokens

Illustrative token counts for common agent tasks, raw file-crawl vs. a graph query.

How OpenVisio builds the map

OpenVisio parses your repository with tree-sitter into a graph of files, symbols, and resolved imports, then ranks what matters with PageRank — no LLM, fully deterministic, and entirely on your machine. It is local-first and read-only: nothing leaves your computer. The result is a map your agent can query over MCP, a visual graph you can explore on the web, and a narrator that explains the code and cites the real path:line behind every answer.

Works with Claude Code, Cursor, and any MCP client

OpenVisio runs as an MCP server, so any compatible agent can resolve task context, find a symbol, trace who depends on a file, or pull a ranked repo skeleton — structured queries instead of blind file reads. The agent gets the context it needs; you stop paying for the context it doesn't.

Frequently asked questions