Back to Tool
S

SelfBench

Tool Oct 17, 2026

An MIT-licensed coding-agent evaluation platform that builds reviewable tasks from merged pull requests and compares model and harness settings by accuracy and cost.

Visit primary link 2 sources attached
SelfBench's author-published interface showing an accuracy-versus-cost chart for a repository benchmark.
Still from the English README's author-published demonstration. Displayed scores, task counts and relative dates belong to that recorded example; they are not current measurements, independent AITrending results or evidence of the timeout-scoring change. Full frame preserved. Source: Mupt AI / SelfBench contributors
agent-ecosystemcoding-agentsai-agents

What happened lately

SelfBench timeline

SelfBench turns merged repository pull requests into coding-agent tasks with instructions, hidden tests and reference solutions. Users review generated tasks before comparing model and harness configurations in sandboxes. Public repository results can be released as accuracy-versus-cost leaderboards.

The web application uses GitHub sign-in and workspace credentials for model and sandbox access. The source is MIT licensed. Local development requires Bun and Docker Compose; the documented reference production deployment uses GCP infrastructure. Repository access and evaluation credentials are operational requirements, not evidence that a benchmark is representative of every software project.