For Teams
Cursor Team Usage Analytics: Dashboard, API & AI Code Tracking
Team and Enterprise admins get a web dashboard with usage analytics: filter by user, group and date (up to 90 days), export CSVs and on Enterprise call the Admin API. The dashboard tracks AI share of committed code (on-device diff signatures), Agent and Tab activity, daily usage, leaderboards, repository insights and, on Enterprise, Conversation Insights and Cloud Agent usage.
On this page
Who can see team analytics?
Team admins see data for themselves and every member. Non-admin members see their own usage and, in some views like the Usage Leaderboard, select peers. Analytics require client version 1.5+. Each chart supports CSV export; the page header can download all charts at once. Enterprise customers can also use the Admin API for programmatic access.
Settle who can see whom before you show anyone the dashboard. Peer visibility makes the leaderboard a social object as much as a measurement, so say what it is for while people are still curious, rather than after somebody has found their own row and drawn their own conclusions.
The failure mode here is using any of it in a performance conversation. Acceptance rate measures a habit and a codebase: an engineer working in generated protobufs will score worse than someone building forms, with no difference in skill between them. Say plainly that the dashboard is not review input, and mean it, because a team that suspects otherwise starts optimising the chart.
This is covered hands-on in Teams and Enterprise Admin — 6 short modules, free to read.
How does Cursor track AI-written code in git?
Cursor logs signatures of AI-suggested lines (Tab or Agent) on-device during chat sessions. Later commits from the same author on the same machine are compared so Cursor can attribute lines written by Tab or Agent versus other edits. Detection stays on-device; only line-count metadata is stored for the dashboard and API.
Automated formatters can invalidate diff signatures. AI code tracking is not yet available for Background Agents or the Cursor CLICursor's command line: the full agent, all modes and models, in the terminal and pipeable into scripts and CI. Press Enter for the full definition.. Commits must be scored on the machine where the AI code was authored.
Read the share as a direction, not an audit. Attribution is computed on-device and tied to the author's machine, so what the chart really reports is code written with Cursor where Cursor could see it happening, which is close to but not identical with the share the AI wrote.
Apply the production-branch filter before you quote the figure to anyone outside the team. The unfiltered view is not limited to what shipped, and a number that includes throwaway branches is the kind of thing a skeptical reader takes apart in one question.
What do the main dashboard charts show?
The dashboard breaks usage into a handful of charts, each answering a different adoption question: how much committed code the AI wrote, where edits come from, who the heavy users are and which repos see the most AI activity. Here is what each one measures.
Map Overview, Members, then Billing
0:39 · narratedRead this demo as text
- cursor.com slash dashboard is the Teams map. Overview shows credits, the setup checklist, and AI Share versus Agent Edits — know the shape before you invite more seats.
- Members is the invite and role job. Who can spend team credits lives here — not in desktop Settings.
- Billing and Invoices covers the Team plan. SSO and SCIM stay behind Explore Enterprise — that banner is a plan gate, not an identity dashboard to invent.
Practice next: Practice this yourself in the hands-on module.
Simulated Cursor 3.12 (macOS, light) — beta educational reconstruction, not the real product.
- Chart
- AI share of committed code
- What it measures
- % of committed lines attributed to Cursor TabCursor's original autocomplete: multi-line, edit-aware suggestions you accept with the Tab key. Press Enter for the full definition. or Agent (optional production-branch filter)
- Chart
- Agent edits
- What it measures
- Code edited by Agent and ⌘K: suggested vs accepted, by extension
- Chart
- Tab completions
- What it measures
- Tab suggestions shown and accepted
- Chart
- Messages sent
- What it measures
- Chat messages by mode (Agent, Ask, ⌘K) and model
- Chart
- Active users
- What it measures
- Unique users per product (Tab, Agent, Background Agent, CLI)
- Chart
- Daily usage
- What it measures
- 365-day activity: all AI edits, Tab, Agent or DAU
- Chart
- Usage leaderboard
- What it measures
- Top users with chats, Tab accepts and Agent lines accepted
- Chart
- Repository insights
- What it measures
- AI vs total lines committed per repo
| Chart | What it measures |
|---|---|
| AI share of committed code | % of committed lines attributed to Cursor TabCursor's original autocomplete: multi-line, edit-aware suggestions you accept with the Tab key. Press Enter for the full definition. or Agent (optional production-branch filter) |
| Agent edits | Code edited by Agent and ⌘K: suggested vs accepted, by extension |
| Tab completions | Tab suggestions shown and accepted |
| Messages sent | Chat messages by mode (Agent, Ask, ⌘K) and model |
| Active users | Unique users per product (Tab, Agent, Background Agent, CLI) |
| Daily usage | 365-day activity: all AI edits, Tab, Agent or DAU |
| Usage leaderboard | Top users with chats, Tab accepts and Agent lines accepted |
| Repository insights | AI vs total lines committed per repo |
Most of these charts are a monthly read at most, and a couple reward a weekly one. Suggested against accepted, on both Agent edits and Tab completions, is the pair that moves when something is genuinely wrong: a steady stream of suggestions with a falling accept rate usually points at prompting or context rather than the model. Active users per product answers a blunter question, which is whether a surface landed at all. Messages sent is the easiest chart on the page to over-read, since a chatty week and a productive week look the same on it.
What is extra on Enterprise?
Conversation InsightsA Cursor analytics view that passively categorises what agents are doing (new features, bug fixes, refactors) so leaders can see where engineering time goes. Press Enter for the full definition. (Enterprise, on by default) classifies agent sessions on-device into work categories (bug fixing, new features, refactoring, DevOps and more) so leaders see what kind of engineering work agents handle without ticket surveys. Compare views adoption across teams and developers. Cloud Agent sections report agents created, PRs opened/merged and lines merged. Conversation Insights inference was free during preview, with pricing applying from January 1, 2026 per Cursor docs.
Conversation InsightsA Cursor analytics view that passively categorises what agents are doing (new features, bug fixes, refactors) so leaders can see where engineering time goes. Press Enter for the full definition. is the feature better raised by you than discovered by someone else. It is on by default, and the classification happens on-device, so the accurate line for a privacy conversation is that the session is read locally and the dashboard receives categories.
That pricing date is the other one, and it has already passed. Check the current terms before you put a number in a budget, because billing mechanics for a preview feature are exactly the detail that moves between quarterly refreshes of a page like this one.
What is the full admin-analytics catalog and how do I read it for adoption?
The admin analytics catalog goes well past the leaderboards. It spans who is using Cursor, how they use it, and on which models and surfaces. The admin's job is to read those signals together to find power users worth cloning and stalls worth fixing.
The catalog: what an admin can actually measure
The reads below are ones you assemble from the dashboard charts and the CSV or Admin API export, not a separate set of first-party metrics. They group the same usage data by the questions worth asking after rollout, so you can move from a raw number to a decision about who to clone and where adoption stalled. Check current chart names against the analytics docs before you wire any of this into a report.
- Metric family
- Usage by surface
- What it reports
- Volume split across Tab, Agent, ⌘K, Ask, Background Agent and CLI
- What it tells you
- Which surfaces have landed and which are dormant
- Metric family
- Usage by model
- What it reports
- Requests and accepted output per model
- What it tells you
- Whether the team defaults to one model or spreads across the catalog
- Metric family
- Work classifier
- What it reports
- On-device classification of agent sessions into bug fixing, new features, refactoring, DevOps, tests and docs
- What it tells you
- What kind of engineering work agents actually carry
- Metric family
- Mode breakdown
- What it reports
- Messages split by Agent vs Ask vs ⌘K
- What it tells you
- Whether people run agents or stay in single-shot ask mode
- Metric family
- Plan adoption
- What it reports
- Active vs licensed seats over time
- What it tells you
- How many paid seats are dormant
- Metric family
- Prompt-specificity score
- What it reports
- A grade on how detailed and context-rich prompts are
- What it tells you
- Whether weak prompting is capping acceptance
- Metric family
- Skills / Hooks / MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. invocations
- What it reports
- Counts of skill runs, hook fires and MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. tool calls per user and team
- What it tells you
- Whether central rules, skills and MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. servers are reaching real workflows
| Metric family | What it reports | What it tells you |
|---|---|---|
| Usage by surface | Volume split across Tab, Agent, ⌘K, Ask, Background Agent and CLI | Which surfaces have landed and which are dormant |
| Usage by model | Requests and accepted output per model | Whether the team defaults to one model or spreads across the catalog |
| Work classifier | On-device classification of agent sessions into bug fixing, new features, refactoring, DevOps, tests and docs | What kind of engineering work agents actually carry |
| Mode breakdown | Messages split by Agent vs Ask vs ⌘K | Whether people run agents or stay in single-shot ask mode |
| Plan adoption | Active vs licensed seats over time | How many paid seats are dormant |
| Prompt-specificity score | A grade on how detailed and context-rich prompts are | Whether weak prompting is capping acceptance |
| Skills / Hooks / MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. invocations | Counts of skill runs, hook fires and MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. tool calls per user and team | Whether central rules, skills and MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. servers are reaching real workflows |
Each family exports to CSV; Enterprise reads the same data through the Admin API.
The read that matters: request volume against acceptance rate
Plot request volume (how much someone leans on Cursor) against acceptance rate (how often suggestions land) and the team sorts itself into four quadrants. The quadrant tells you who to clone and who to coach in one glance: top-right are power users, bottom-right are people fighting the tool.
Power users. Read their prompt-specificity scores and MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition./skill invocation counts, then turn what they do into shared rules and skills.
Trying hard, getting rejected. Usually a prompting or context problem. Check the specificity score and pair them with a power user.
Cautious but effective. The suggestions they take are good; the opportunity is breadth. Nudge them onto Agent and more surfaces.
Stalled or dormant seats. Cross-check plan adoption: a dormant paid seat is a chargeback you can reclaim or a person who needs onboarding.
Adoption is a loop, not a report. Push rules and skills to the team centrally, then watch the Skills / Hooks / MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. invocation counts climb. If invocations stay flat after you publish a skill, that's a discovery gap. Surface it in onboarding and pair it with a power user.
Re-read the prompt-specificity score after you ship prompting guidance. A rising score with a rising acceptance rate is the clearest signal your enablement is working.
Frequently asked questions
Can we export analytics data?
Yes. Each chart has a CSV download; the dashboard header can export all charts. Enterprise can also use the Admin API.
Does AI code tracking send our source code to Cursor?
No. Signatures and line counts are computed on-device; only metadata line counts are stored for reporting.
How far back can we filter analytics?
Up to 90 continuous days, with filters for up to 10 users and optional timezone and weekend display in settings.
Sources & last verified
Cursor ships frequently. Last updated July 28, 2026.
Keep reading
Rather do it than read about it? Run 11 interactive Cursor walkthroughs in a simulated editor. Free, no account needed.