AI impact

DevGlimpse shows what AI-assisted and agent-authored work actually ships, next to human work. It uses two sources. Git history tells you which commits and pull requests AI wrote or helped with. Usage telemetry from the AI tools themselves tells you who uses them, how much and at what cost. Use it to learn where AI helps and to share good practice, not to rank people.

Git-based AI reports

You need no setup for these. DevGlimpse reads your connected repositories.

AI Tool Contributions (AI Impact category) classifies work from signatures in git:

• A commit counts as AI-authored when its author, or a Co-authored-by trailer, is a known AI tool such as Claude Code, Codex, GitHub Copilot, Cursor or Devin, or another listed agent.
• A merged pull request counts as AI-assisted when it has an AI-authored commit, an AI bot as its opener, or an agent marker in its description.

The report shows:

• AI-authored and human-only commits, the AI share of commits, and a week-by-week stacked chart.
• AI-authored commits per person, with search. Select a person to open their page.
• AI-assisted against human pull requests on cycle time, size and review latency.
• AI quality: how often each kind was later reverted, how closely each was reviewed before merge, and how often each was linked to a later bug or incident.

Add your own AI signatures

If your company runs an internal agent, teach DevGlimpse to recognize it.

1. Open Admin → AI signatures. Managers can read it. Only an org admin can change it.
2. Under This workspace’s signatures, choose the Tool, the Signal and a Pattern.
3. Choose a signal: Commit author login, Commit author email, Co-authored-by trailer, PR opener login or PR body marker. The page shows an example for each, such as house-agent[bot].
4. Add the signature. DevGlimpse reclassifies your existing history, so the reports update once that run finishes.

The Built-in catalogue on the same page lists the signatures DevGlimpse already knows. It is read-only.

Connect Claude Code, Codex CLI and Gemini CLI

The tools send usage events to DevGlimpse over OpenTelemetry. Only an org admin can set this up.

1. Open Admin → AI sources. You can also use Start on Add AI Source in the Setup tasks card on Home.

2. Under Ingest keys, type a Key name, such as Platform team laptops, and select Create key. Create one key per team or group of machines so you can revoke it on its own.

3. Copy the key now. It is shown once and cannot be retrieved. Select I’ve saved it when done. If you lose it, revoke it and create another.

4. Under Configure your tools, open the tab for your tool. The snippets already contain your new key.

   • Claude Code: add the lines to the shell profile of each developer, then restart Claude Code.
   • Claude Code (managed): deploy the settings file to every managed device with your device management tool. The page lists the path for macOS, Linux and WSL, and Windows.
   • Codex CLI: add the block to ~/.codex/config.toml, then restart Codex. Codex ignores this block in a project-local file.
   • Gemini CLI: merge the telemetry block into ~/.gemini/settings.json, add the line to ~/.gemini/.env, then restart Gemini CLI.

5. Start a session on a configured machine. The Events received card changes from Waiting for events to Receiving and shows the event count and the time of the latest event. It checks every few seconds.

Match telemetry to people

Telemetry counts for a person only when its identity maps to a contributor.

1. On Admin → AI sources, look at Unmatched identities. It lists emails and account or user ids that match no contributor, with the tool, the number of events and the last time seen.
2. For each, choose the contributor it belongs to under Map to, or ignore it. Its stored events then count for that person, or never count.

Gemini CLI sessions are attributed to a contributor only when the developer signs in with Google. Sessions that use a Gemini API key or Vertex AI arrive unattributed.

Adoption, usage and spend

AI Adoption (AI Impact category) reads the tools’ own telemetry:

• AI output vs. merged code: per tool, the lines added and removed, commits and pull requests the tool reports creating, edits accepted and rejected, the acceptance rate, and the AI-authored commits found in git. Codex reports none of the line and commit counts, and Gemini CLI reports only lines.
• Active users by tool and by team, week by week.
• Tokens and cost by tool and model, by team and by person. Claude Code reports a cost estimate in US dollars. Codex and Gemini CLI report tokens only, so cost shows a dash.
• Human intervention and session automation: how often a person steps in while a tool works, and how much the tool does between those steps. Tools that send no such events show Not available.
• AI adoption date and ROI: for each team, the first week AI-assisted pull requests passed a share of merged pull requests, or a date an org admin sets by hand. The report compares a window before that date with an equal window from it (the window is shown on the chart), and says it shows correlation, not causation.
• AI spend by ticket and initiative: each session’s spend, followed to the pull requests it contributed to, then to their linked tickets and initiatives. Parts with no link are shown apart, so the parts add up to the total.

AI Sessions shows how long sessions run and how many tool calls each makes, week by week and by team. Longer is not better or worse by itself.

Team lines are hidden while a team is smaller than the workspace’s minimum group size. People appear only if you may see their individual metrics.

Your own private AI view

Open My AI usage (Home category) to see your own sessions, how long they run, how many tool calls they make, and the tokens and cost they used. Nobody else’s usage is shown there and nobody is ranked. If you see No contributor is linked to your GitHub login yet, make sure a commit, pull request or review of yours has synced, and that your telemetry identity is mapped to you.

Common questions

Related

• More AI vendors
• Privacy and control
• People and identities
• Metric glossary