See the spend. See the work. Scale what ships. Open-source analytics across Claude Code, Codex and the rest of your dev-AI stack.
One view of your team's spend, sessions, and outcomes across Claude Code, Codex, and Cursor — in a single dashboard that reads session files directly from each dev's machine.
What your agents cost, what they shipped, and which prompts actually ship code. Built for engineering leaders handed an AI bill, a pile of session logs, and asked to make sense of both.
The collector auto-detects coding agents on the engineer's machine, parses native session files locally, and normalizes spend across models. Pricing pinned at capture, so model-price shifts don't silently rewrite history.
The GitHub App joins sessions to merged PRs through accepted-edit events, AI-Assisted commit trailers, and a git-log fallback. Cost per merged PR with the commits that earned it. Sessions that shipped vs sessions that burned.
Cluster the team's prompts, find the cohort that solved the same problem cheaper, and promote winning workflows as playbooks. The pattern becomes the asset. (Vision — backend live, dashboard gesture lands next.)
| Adapter | Status | What it captures |
|---|---|---|
| Claude Code | Full | Sessions, input/output/cache tokens, models, tool calls, accepted edits |
| Codex | Full | Sessions, per-turn token diffs, tool executions, dollar cost |
| Cursor | In dev | Messages, lines suggested, accept rate. Subscription-billed — no per-request cost exposed. |
| Continue.dev | In dev | Chat turns, token generation, edit outcomes, tool usage |
| OpenCode | In dev | Sessions, tokens, model routing (SQLite, post-v1.2) |
| VS Code (generic SDK) | In dev | Pluggable handlers via SDK — community adapters supported |
Sessions that end in merged code.
Accepted edits per dollar, peer-normalized.
How often a session ships without a hand-hold.
How many of your agents the engineer actually uses.
Playbooks this engineer promoted that others adopted.
The most expensive system your engineering org has ever bought may be the one you understand the least.
Bema measures it. Spend across every agent. Outcomes tied to merged code. The prompts that ship, surfaced and shareable. Open-source, self-hostable, runs against your local sessions on day one. The data was always yours — now it's an instrument.