Across six questions about Hono, a 451-file open-source framework, Nebulai lowered measured session cost and token use while giving the agent relevant Project Context.
lower, across two repeated runs of the same question set
fewer, same runs
First run: 13 of 17 successful pairs
Savings compare attempts where both agents passed the recorded grader: 17 pairs covering six questions in the first run, and 15 pairs covering five in the second. Recorded success was 100% versus 94.4%, then 94.4% versus 83.3% (baseline versus Nebulai). All failed attempts missed one literal naming check; factual correctness needs separate review. These figures exclude memory construction and hosting costs and do not measure cross-session discovery reuse.
| Question | Rounds unaided | Saved |
|---|---|---|
| context-response | 8.0 | +34.3% |
| error-propagation | 9.0 | +29.5% |
| middleware-composition | 10.0 | +18.6% |
| request-params | 15.3 | +17.4% |
| smart-router-selection | 6.0 | +5.9% |
| regexp-route-compilation | 7.0 | +3.6% |
Nebulai helped the agent reach the relevant parts of the codebase with fewer tokens and lower measured-session cost across both benchmark runs.
That is the outcome Nebulai is built for: engineering teams and coding agents start with the project’s existing understanding instead of paying to reconstruct it.
Nebulai is currently in private beta. We’re opening early access to teams that want their software projects to remember.
Request early access