How unerr cuts code-navigation tokens 86–90%

Your coding agent spends most of its tokens reading your code, not writing it. Every task starts with the same ritual — list files, open files, grep, re-open the same module it saw ten minutes ago. That navigation-and-read loop is the dominant token cost of agentic coding, and it's what unerr attacks.

This post is the methodology and the numbers. The short version:

Tool	Navigation tokens	Read tokens
unerr	−86%	−90%
graphify	−43%	−81%
RTK	−30%	−49%

Same repos, same tasks, same tokenizer, head to head.

What we measured

The benchmark replays a fixed set of code-navigation and code-reading workloads — the calls an agent actually makes while working: find a symbol, trace its callers, read the relevant slice of a file, get the structure of a module. Each workload runs twice:

1. Baseline — the agent's stock tools (full-file reads, grep-style search).
2. Tool under test — the same information need served by unerr, graphify, or RTK.

We count the tokens the model would ingest in each case, with a real tokenizer — not character-count approximations.

The fidelity gate

Compression is trivial if you're allowed to drop information. Every result above is fidelity-gated: the compressed output must still contain the facts the workload needs (the symbol, its signature, its callers, the code slice). A response that saves tokens but loses the answer scores zero. The percentages are savings at equal usefulness, not savings at any cost.

Why the gap

unerr serves agents from a local code graph built ahead of time — entities, references, structure — so a "who calls this?" question is a graph lookup, not a repo-wide grep plus a stack of full-file reads. Reads return the relevant slice with structure, not the whole file. The graph answers in under 5 ms locally, supports 8 languages, and ships 22 MCP tools that agents (Claude Code, Cursor, Copilot, Windsurf, and any MCP-compatible agent) call natively.

Scope — what this number is and isn't

The −86%/−90% is measured on navigation and read workloads — the dominant cost, but not your whole bill. We publish 75%+ as the conservative public figure for overall savings, and the per-agent translation (longer sessions on Claude Code, credits going further on Cursor, smaller API bills on BYO-key agents) varies with each agent's billing model. We'd rather you verify than trust:

Reproduce it on your repo

$ tsx benchmarks/agent-token-bench/run.ts <your-repo>

The harness clones nothing and uploads nothing — it runs locally against your working copy, prints the per-workload table, and the fidelity gate is in the open. If your numbers differ from ours, we want to know.

Benchmark run against unerr v0.2.11.