SOFTWARE ENGINEERING

CodeMesh: 95% fewer tokens for AI coding agents

When AI coding assistants moved to token billing, an engineering leader at a global energy-equipment manufacturer asked how to cut the cost. Wexa built CodeMesh: a structural graph of the codebase that coding agents query over MCP instead of reading whole files.

Download this case as PDF Book a demo

fewer tokens across 98 engineering queries (25.3M → 1.23M)
fewer tokens per query (258,115 → 12,524)
measured API spend for the benchmark
savings held on hard and expert queries

CLIENT: Wexa product, built from a problem raised by a global energy-equipment manufacturer
REGION: Global
WHEN: 2026
ENGAGEMENT: Wexa product
WEXA COMPONENTS: CodeMesh on the Wexa memory layer · MCP server · 5 retrieval tools

01Case 10

The challenge

Coding agents re-read whole files and rebuild context for every query
Token-based billing makes that cost visible and large
Agents miss relationships between functions, classes and dependencies

What Wexa built

Code knowledge graph of files, functions, classes, imports and git history on the Wexa memory layer
Exposed over MCP to Claude Code, Cursor, ChatGPT and custom agents
Guard hook redirects file search to the graph; no LLM in the retrieval path
Read-only, permission-checked access with audit of every agent query

02Case 10

How it runs on Wexa

FigureHow this use case runs on Wexa: client systems connected, ontology discovered, context graph in the Wexa memory layer, agents and process flows, and governance. Highlighted components are central to this use case.

03Case 10

Solution architecture

FigureCodeMesh architecture: development context → CodeMesh → structural graph → AI coding agents.

04Case 10

How the result was measured

FigureBenchmark method: one public codebase, 98 real engineering questions, the same model asked twice, every token counted.

05Case 10

Results

Measure	Before / without Wexa	With Wexa
Coding-agent token spend (team of 50 developers)	$300K a year	$21K a year
Annual cost of the process	$300K	$21K
Time to implement the AI solution	4 months	1 day
AI build cost (one-time)	$192K	$4K
AI operating cost per year (models, infrastructure, upkeep, platform)	$141K	$66K
Total AI cost in year 1	$333K	$70K

With Wexa

net annual value to the client after Wexa's own operating cost

faster to implement

lower AI operating cost

less AI spend in year 1

Public benchmark on an open-source repository; method and questions published for anyone to reproduce
On a private production repository, a 10-question test measured 82–90% input-token reduction
Prospects include a global IT-services firm and a US hospital group

Benchmark source: CodeMesh benchmark (codemesh.cloud/benchmarks). Same model and questions in both runs.