Use it
CLI
Every aa command.
The binary is agentic-analytics, alias aa (bun run aa … from a clone).
| Command | Does |
|---|---|
aa analyze <paths…> [--yes] [--objective "…"] [--web] [--small-model] [--out dir] [--open] [--currency CODE] [--embedder hash] | A full job with a live stage view and, without --yes, an interactive review of the drafted metrics (accept, edit as YAML, drop). Writes <out>/agentic-analytics/report.json and report.html. |
aa ask <job> "question" | A follow-up answer: claims that passed the verifier, with numbers from the report’s facts. |
aa explain <job> <fact> | The computation behind a number: SQL and parameters (or function and inputs), its definition, and the input rows. |
aa serve [--port 8080] [--host 127.0.0.1] [--allow-no-auth] | The HTTP API, with the viewer at / and MCP at /mcp. A non-loopback --host needs API_TOKEN (or --allow-no-auth behind a proxy or a localhost-only port mapping). AA_MCP_ROOTS lists the folders MCP over HTTP may read. |
aa mcp | The MCP server on stdio. |
aa jobs ls | rm <id> | resume <id> | Manage jobs in AA_JOBS_DIR (default data/jobs). |
aa eval [--suite smoke|full|heldout|real] [--model …] | The eval suite; updates evals/RESULTS.md. With --suite heldout it runs the held-out generated set, for release runs. |
aa eval --suite full --label "<name>" [--note "<text>"] [--only <id,…> --results-dir <dir>] | A run under its own model row (a self-hosted model); --only runs some businesses, joined later with bun run eval:merge. |
aa eval --suite full --model configured --passes 2 | A run over up to two nights (what scripts/eval-daily.sh runs): the next night resumes it and runs only the businesses a free tier’s quota stopped. It is partial only if businesses are unfinished after the last pass. |
aa eval --suite smoke --repeat 5 [--model <m>] | Consistency: the suite five times with the response cache off, then the share of businesses where every run agrees on the numbers, the top findings, each finding’s driver and cause grade, and the top action. The disagreements are written side by side in evals/results/repeat/. |
aa eval … --no-cache | Every model call live: nothing read from or written to the response cache. --repeat implies it. |
aa eval --suite real --answers <dir> [--data <dir>] | The real-data checks on your own files (tutorial); results in ./eval-results. |
aa usage [--day YYYY-MM-DD] [--log <path>] [--json] | Tokens logged per route in the last 24 hours, with the provider’s own count of its daily limit when a rate-limit error reported one, then tokens and calls per day (UTC), per role and per route, from data/logs/model-calls.jsonl, against each free tier’s daily limit where it is known. Cached calls count 0. |
aa doctor | What this machine has, with a fix for each missing piece. |
aa metrics <paths…> | Draft and review metric definitions only; --save writes them for reuse. |
aa run <paths…> --dir <folder> / --resume <folder> | A job in an explicit folder, for scripts. |
Configuration comes from .env and agentic-analytics.config.yaml: see Config.