Track ({ coding agents }) Discover ({ problems })
// Learn, improve

We store it so you can search it later - that’s the product, not a side effect. Accounts are individual: nobody sees your sessions unless you join a club or team and choose to share. We don’t train models on your data, and neither do the providers we use. Before anything uploads, bmor strips it on your machine - file contents the agent read, command output, the bodies and diffs from every Write and Edit, pasted screenshots, extended thinking - then scans what’s left against 200+ known secret formats (AWS keys, GitHub tokens, private keys, JWTs) and redacts matches in place. Encrypted in transit and at rest in AWS us-east-1. Deleted within a month if you ask. More in our Security overview.

Token counts tell you what a session cost. The transcript tells you why. It’s the record of what your agent actually did, which is what you want the day a file disappears, a secret gets committed, or a deploy goes sideways. It’s also what makes the numbers actionable: where the model pushed back on you, which tool calls are worth a second look, why one session cost 8× its neighbors, what to fix in your CLAUDE.md or AGENTS.md. And your agent cleans up its own local logs on a rolling window - Claude Code defaults to 30 days. We keep yours.

Three things. Create an account. Install the CLI with one command - curl, Homebrew, or npm. Run bmor init: it writes a default config, detects which agents you already have installed, registers their hooks for you, opens a browser to log in, and asks one question - full transcripts or stats-only. That’s the five minutes. Then run bmor sync backfill and your existing session history uploads, so your dashboards have real data on day one instead of next week. No API keys to paste, no config to hand-edit, nothing to change in your repos. Uninstall is bmor uninstall.

Your agent already writes a session log to disk. bmor reads it. On install it registers hooks in each agent’s own config - ~/.claude/settings.json, ~/.codex/config.toml, ~/.cursor/hooks.json - that fire on session start and session end, plus Claude’s status line, which is where rate-limit data comes from. The hook spawns a detached background process, so your agent never blocks on our network. We don’t proxy your API calls, wrap your agent, or sit anywhere in the request path. If we’re down or you’re on a plane, batches stage locally and ship on the next sync. The agent never notices.

Today: Claude Code, Codex, and Cursor. Two things decide what’s next - how many people ask, and whether the agent leaves a local trail we can read. If it writes session logs to disk, we can usually add it. Cursor doesn’t, so we read its local SQLite database instead. Each new agent is a provider module - parser, collector, redactor - plus the work of normalizing its numbers into the same shape as everything else, so you can compare tools side by side instead of squinting at three dashboards. We don’t publish dates we can’t hit. Tell us what you’re running on Discord or the feedback page and it moves up the list.