yeaboi

Agents Beta

The other half of your team doesn't come to standup. The Agents family answers the three questions nobody else is asking about your AI coding agents: what they cost, what they did, and what they can reach.

Beta. The Agents family is new, and every number it shows is computed from telemetry your agents happen to leave behind. Treat it as an indicator, not an audit.

Two audiences, one command

yeaboi opens on a landing split — Who are we working with today? — with a world-card for each side. Humans opens the scrum modes: planning, standups, retros, poker, reviews. Agents opens this family. The split is always shown on a fresh launch; the side you picked last is preselected, never auto-skipped, so neither family can become invisible. Inside a menu, esc returns to the split and q quits.

The three modes

  • Usage — tokens, cache traffic, and spend per model, per project and per day.
  • Standup — a daily digest of the sessions your agents worked and the commits and PRs they authored.
  • Security — an audit of the setup behind them: permissions, MCP servers, secrets exposure, risky commands.

Where the data comes from

The collector reads the append-only JSONL transcripts Claude Code writes under ~/.claude/projects/ and rolls each file up into one session record: per-model token totals (including the 5-minute and 1-hour cache-write split), tool-use counts, turns, project path, branch and timestamps. Claude Code is currently the only telemetry source.

Two details are worth knowing because they decide whether the numbers are right. First, one assistant API message is split across several JSONL lines that share a requestId and repeat identical usage, so usage is counted once per requestId — counting lines would inflate every total. Second, a file whose size and mtime are unchanged is skipped without being opened; any change re-parses that whole file and its rollup replaces the previous one, so there is no partial-offset double counting.

Malformed lines are counted and skipped, and a file that fails to parse becomes a warning rather than an error — a refresh never raises.

What leaves your machine

Nothing. Every figure on these pages is computed locally from session logs that are already on your disk, and the results are stored in the same local SQLite database the rest of yeaboi uses. There is no yeaboi service to send them to.

The stronger guarantee is about what yeaboi itself keeps. The collector makes exactly one pass over raw transcript text, in the stream, and persists no transcript content at all — not prompts, not code, not the commands an agent ran. A security finding carries a pattern label, a file path and a line number, and never the secret it matched. That boundary is enforced by tests, not by convention.

The one narrow exception is your own configuration: a finding may quote a permission mode or an allow rule from your settings file, because “an allow rule auto-approves bash” is not actionable without naming the rule — and your settings are not session content.

Every surface

Like the rest of yeaboi, the Agents family is not TUI-only. The same three engines back the TUI cards, the yeaboi agents subcommands, the agents_* MCP tools, and the agents-usage, agents-standup and agents-security plugin skills.

yeaboi agents cost       # what they cost
yeaboi agents standup    # what they did
yeaboi agents security   # what they can reach

See the CLI Reference for every flag.