Simple code fix — 18.67%
Median token reduction across three verified OFF/ON pairs. This is the source of the rounded “about 20%” headline.
Controlled Claude Code A/B experiments · Product evidence
RepoMesh token benchmarks: what we measured
In a small, controlled experiment, RepoMesh delivered fewer provider-reported tokens for three verified coding scenarios. The homepage's about 20% fewer tokens claim rounds the 18.67% median reduction from the simple-fix scenario. It is a measured example, not a promise for every task or model.
Results by coding task
Each verified scenario used three matched pairs of RepoMesh OFF and RepoMesh ON runs. Percentages are the median reductions in provider-reported total tokens within that scenario.
Multi-file bug — 34.71%
Median token reduction across three verified pairs. Median agent execution time was also 19.3 seconds lower in the treatment runs.
Dependency regression — 30.62%
Median token reduction across three verified pairs. Median agent execution time was 20.4 seconds lower in the treatment runs.
How the comparison was run
The evaluation used a 34-file TypeScript booking application, Claude Code and Claude Sonnet 4.5. Each paired run was held to the same task, model and objective quality checks.
1. Matched conditions
The OFF control used no RepoMesh context service. The ON treatment had a task-scoped RepoMesh MCP connection. Both were evaluated on the same source snapshot and task.
2. Actual model usage
The numbers use provider-reported input and output tokens, not character-count estimates from RepoMesh's dashboard. Cached input is not counted twice.
3. Verified delivery and quality
A reported pair required a successful task, objective acceptance checks and observed RepoMesh context delivery in the ON run. Unverified runs were excluded, not assigned zero savings.
4. Three pairs per scenario
We use the median of the three verified paired reductions for each scenario. The percentages are not aggregated into one universal savings figure.
What these results do not prove
Three matched pairs per scenario are not statistically significant. The benchmark is tool-assisted: both sides were instructed to use a repository-context service when available. It does not prove that an agent will discover RepoMesh by itself, or that every task will consume fewer tokens.
One additional feature-test scenario produced no verified result because its acceptance criterion required an untouched file. It is not included in the figures. Median time differences were measured in two scenarios, but they do not establish a general speedup. Token reductions are not necessarily equivalent to Claude Pro subscription savings or a lower invoice.
Results may differ with the repository, task, agent, model, prompt, cache behavior and test environment. These are RepoMesh's own controlled measurements, not an independent third-party study.
Frequently asked questions
Does RepoMesh guarantee 20% fewer tokens?
No. About 20% is a rounded observed median for one simple-fix scenario. Two other verified scenarios had larger median reductions; no fixed savings are guaranteed.
What does RepoMesh do during a coding task?
It indexes repository structure and supplies bounded, task-relevant files, symbols, dependencies and tests through MCP. The connected coding agent still decides what to inspect, edit and run.
Is RepoMesh itself an AI coding agent?
No. It is a local-first repository intelligence layer. It works with compatible clients such as Claude Code, OpenAI Codex, Cursor and OpenCode; it does not generate application code or route models.
Where can I learn how it works?
Read the MCP repository context guide, the supported agent integrations and the GitHub Remote overview.