Team and developer analytics for AI coding agents

The Developers space in Caveman Cloud turns personal-key traffic into team analytics: who is using which coding agents, what it costs at catalog prices, which sessions drove the spend, and which merged changes those sessions led to. Every view is scoped by the reporting window and project pickers at the top of the page. Set up personal keys first with Coding agent analytics.

​

Team

Developers → Team treats each project as a team. For sub-teams, create more projects.

• Active people: people with attributed requests in the window, including observed contributors.
• Requests: model calls on personal keys attributed to verified people. A request is a model call, not finished work.
• Catalog spend: personal-key spend at catalog list prices, not the provider invoice.
• Cache read: input served from the provider cache. This is reuse, not a verified saving.

Panels below break this down by Requests by person, Coding agents, Attribution, and Seats used. Attribution shows the share of catalog-priced project spend that is linked to people. The rest is app traffic or keys no person holds.

​

Individual

Developers → Individual shows one person’s activity. Pick an assigned member, an observed contributor, or a former member.

• Tiles: Requests, Catalog spend, Coding agents, Cache read, Share of team, and Active days (UTC days with at least one request, not time worked).
• Panels: Coding agents, Models (by catalog spend), Work types (the label the router classifier inferred), and Daily requests.

Developers can see their own data any time in My usage, including their keys and machines.

​

Agent usage

Developers → Agent usage compares coding agents side by side.

• Tiles: Coding agents, Requests, Catalog spend, and Cost / request. Cost per request moves with the model and task mix, so a drop is not proof of a saving.
• Cache read by work type shows where reuse happens.
• Coding agents compared lists each agent’s requests and per-request cost, with View traces to drill into the calls.

​

Your agents

Developers → Your agents shows machines that run the caveman CLI wrapper. Use Whose agents to switch between people. Tiles show Machines, Requests, Measured tokens, and Kept out of context, and the Machines and agents table breaks down routing, model mix, and measured tokens per machine. Machines appear after you run caveman login.

​

Sessions

Developers → Sessions lists coding sessions in the window. A session appears once it runs through the gateway with a personal key.

• Sort by Recent activity or Highest cost, choose columns, and switch between Everyone and Only mine.
• Search by ID, repository, or model, and add filters.
• Tiles show Sessions, Requests (errors are HTTP status 400 or higher), Priced cost, and Cache read.
• Peak context is the largest provider-reported input token count in the session, not a percentage of the context window.

Open a session to see its timeline and inspect individual requests.

​

Repositories

Developers → Repositories joins GitHub pull requests to agent sessions. Once connected, tiles show Merged changes, Linked to usage, Open to merge (median time from opening a pull request to merging it), and Linked spend. A change is linked when its branch matches session usage in the same project and repository. Linking does not say which lines an agent wrote. Below the tiles, the Time to merge chart, a per-repository list, and the Merged changes table show each change’s open-to-merge time, sessions, and evidence. Sort by Recently merged or Longest open-to-merge, and search by title, branch, or GitHub author.

​

Delivery

Developers → Delivery follows the money from session spend to outcomes. It needs GitHub connected and sessions that carry repository and branch metadata.

• Tiles: Session spend, Linked merges, Median linked cost, and Branch context.
• Where the spend went traces spend through branch context to an outcome: Still moving, No merge observed, or Needs evidence.
• Branches summarizes branches that are still moving or quiet with no merge, plus P90 linked cost and median sessions.
• Change size and Merged work group merged changes by size and kind (read from pull request and branch names).
• Quiet branches lists branches with no activity for 48 hours and no merge. A quiet branch is not automatically waste.
• Optimizer cohorts compare observed outcomes with and without optimizers. These are observations, not causal results.

If sessions are missing work context, send repo=owner/name,branch=<branch> in x-cave-tags together with an x-cave-session ID.

​

Who can see what

Viewers without billing access see Active people, Requests, Attributed, and Members, but no money values. Grant billing access in Access and limits.

​

Next steps

• Chart coding agent spend on a custom dashboard
• Set budgets per person or key
• Allocate spend to teams and cost centers