Skip to main content
AI-native metrics are part of the Pro plan. See Plans.
As teams adopt coding agents (Claude Code, Cursor, Copilot, and others), the important questions change: how much work is AI-assisted, does AI-written code get accepted and stay, and does it shift the review burden? DevPerform measures all of this from your existing code-host data (GitHub, GitLab, or Bitbucket) — no IDE plugin required.

​
How authorship is detected

DevPerform reads commit metadata — message trailers (“Co-authored-by”, “Generated with” lines) plus the git author identity that cloud agents commit under — to attribute work to a tool: Claude Code, Cursor, GitHub Copilot (coding agent, VS Code, CLI), Devin, Aider, OpenAI Codex, Windsurf, Google Jules, Gemini Code Assist, OpenHands, Amp, Amazon Q, Kiro, JetBrains Junie, Replit Agent, Factory Droid, and Ona. Each PR is then classified three ways:
These numbers are a lower bound. AI use that leaves no trace in commit metadata is invisible to detection, so real AI involvement is at least what’s reported, often more. DevPerform states this honestly rather than overclaiming.

​
The metrics

Provenance

Share of merged PRs and commits attributed to AI, broken down by tool and by human / assisted / agent.

Agent acceptance

How often AI-attributed PRs get merged vs. closed unmerged, compared to human PRs.

Rework on AI code

Whether AI-written lines get rewritten more or less than human-written lines (built on the rework engine).

Review-burden shift

Comments per PR, time-to-first-review, and reviewed rate for AI vs. human work — does AI move the bottleneck to review?

Revert rate

How often AI-authored commits get reverted compared to human commits — a signal of whether AI code stays once it ships.
Most of these also expose a human / assisted / agent split, since agent-authored work is usually the sharpest cohort to watch.

​
How the data fills in

AI attribution is stamped as commits are synced, so the numbers backfill gradually — only re-synced commits carry authorship. Expect coverage to improve over the first sync cycles after enabling Pro.

​
AI Impact views

The AI Impact page (Pro) adds two team-level analyses:
  • Adoption cohorts — with four or more teams, DevPerform splits them into higher- and lower-AI-adoption cohorts at the median and compares cycle time, throughput per contributor, unreviewed-merge rate, and revert rate. Below four teams you get honest per-team rows instead of a forced split.
  • Adoption vs delivery trends — weekly AI-authored PR share charted against cycle time, throughput, and revert rate, so “did delivery change as AI ramped?” is answerable at a glance.
Both views measure adoption from AI-attributed commits (a lower bound) and carry a correlation-is-not-causation note, plus an industry context line citing openly published adoption figures (Stack Overflow’s 2025 developer survey and Google’s DORA 2025 report, linked on the page) so your adoption number has a reference point. DevPerform never tracks individual AI usage — the smallest unit in any AI metric is a team.

​
AI Tools comparison

The AI Tools page (Pro) compares detected AI tools side by side: attributed PRs, commits and lines, revert and unreviewed-merge rates per tool, and a weekly adoption chart. Enter each tool’s monthly spend (kept in your browser — never stored server-side) to get cost per AI-assisted PR by tool and a spend-share vs work-share comparison. Detection is commit-signature based — message trailers plus the git author identity that cloud agents commit under — a lower bound whose coverage varies by tool; the page says so. It also shows seat utilization: how many people actively used each tool in the period — as a count only, never a list of names — and, if you enter how many seats you pay for (also browser-only), the idle-seat count and the spend tied up in them. “You pay for 20 Copilot seats and 8 were active this month” is often the fastest AI-cost win there is.