Skip to main content
Use fp to inspect Cloud telemetry, manage cloud-managed enforcement (policies, fleet deployments, guardrail decisions), and manage audits, findings, issues, alerts, keys, users, queries, and settings. Use failproofai for local hooks, policies, capture, and machine enrollment. You can also author and deploy hosted code, LLM-judge and Jev evaluators, manage dashboards, inspect sentiment, and use the dashboard assistant from the terminal. Install the released Cloud CLI as an isolated tool:

Sign in

Syntax

Global options must come before the command:
Run fp COMMAND --help or fp COMMAND SUBCOMMAND --help for terminal help.

CLI commands

Author an evaluator

fp evaluators authors definitions; fp evals reads results. An enabled managed instance is required (fp evaluators instances). Authoring needs evaluations:trigger; session dry-runs and schema profiling also need events:read. Code is one restricted Python expression. Save this as error_free.py:
LLM judge:
Jev binary and graded rubric:
Creation enables the evaluator. Deploy creates a new immutable version, then disables older enabled versions. A failed follow-up leaves the new version published: inspect evaluators list --all before retrying. Judge dry-runs are refused by Cloud. Jev dry-runs spend model budget. Backfill covers code, judge and Jev evaluators and accepts at most 90 days; it queues only sessions an evaluator has not scored yet. Judge and Jev backfills spend model budget, so run fp evaluators backfill --dry-run first. Human output shows per-session dry-run results before creating/deploying. Failed previews stop the write unless you explicitly pass --force after review.
AI drafting requires fp login and writes a spec accepted by create/deploy:
Code flags are --code FILE|-, --result-kind score|metric|assertion, --timeout S. Judge flags are --judge --criteria TEXT, repeated --step and --threshold. Jev flags are --jev noul --instructions --true --false or --jev score --instructions with 3–5 distinct --level values, worst first. Judge/Jev require a score; thresholds are 0–1. Binary judge steps are rejected. Spec files contain exactly eval_key, display_name, eval_version, result_kind, labels, condition_source, evaluator_source, timeout_seconds. Judge/Jev evaluator_source is a JSON envelope string. Do not include row ids, enabled state, or AI reasoning metadata. --file and discrete spec flags are exclusive; deploy can override the file’s version. Output files are never overwritten.

New dashboard controls

All commands accept global --json. Unknown-route 404s exit 6 with “not available on this FailproofAI Cloud yet”. Missing-record 404s also exit 6 but report that the record was not found. Session-only commands deny key mode before HTTP (exit 2). Human commands use summaries and tables. For raw JSON exports, pass global --json or save with --out FILE; otherwise session and JSON query exports show a readable view. CSV exports remain plain CSV for piping. fp agent ask renders artifact SQL as tables/stats through queries:run. Write approvals show the proposed change and prompt on a terminal. Scripts and --json decline unless --approve is explicit. Existing JSON shapes stay unchanged when no artifact is returned.

Authentication

On a terminal, fp login starts on a link, to sign in with your browser or SSO, and opens it (over SSH or with no display it only shows it, and it works from any device); authorize it and the terminal finishes on its own. That terminal gets your access in all your organizations until the session expires, so only authorize a link you just asked for. If the login started on a different network, you get an email once the session exists, and the page warns you where the deployment can tell. Press Tab for a 6-digit code by email instead, and Tab again to go back. --email, --json and anything without a terminal use the code.

Events

Lists individual agent events. The default light feed excludes raw payloads; use --full only for a bounded investigation.
--all paginates up to --limit, which defaults to 50 — so --all on its own stops at 50 rows. When it stops early the response carries a next_cursor to resume from; "next_cursor": null means the feed really was exhausted.

Sessions

Evaluations

Errors

Usage and filter values

Organizations

API keys

Permission tokens use resource:action, such as events:add. Repeat --add, comma-separate tokens, or use dotted actions such as events:read.add.

Queries

Users

Settings

Alerts

Alert severities are info, warning, and critical. Trigger kinds are metric_threshold, custom_sql, evaluation_score, eval_compound, and per_event. Evaluation intervals must be between 30 and 86,400 seconds.

Audits

Audit create options

Include context during creation when the first run needs it. Creation commits the definition and context together before the queued run begins.
fp audits run is asynchronous. Poll fp audits runs NAME until the latest run succeeds or fails before reading its findings.

Issues

Valid issue states are firing, acknowledged, and resolved. Standalone issue severities are info, warning, and critical.

Cloud assistant

Policies

Cloud-managed policy versions. Session-only — every command here exits 2 under an API key, before any request, because these are root-only write routes deliberately absent from /v1.

Fleet

Which machines run which policies. Session-only, same reason as above.

Guardrails

What enforcement actually did. Session-only, same reason as above.

Global flags

--api-key is intended for automation. Login, organization switching, and assistant commands require a user session.

Environment variables

Explicit flags override environment variables, which override saved configuration. In API-key mode, select the tenant explicitly with --org or FP_ORG.
The AGENTEYE_* spellings of these are not read by fp and never were — the CLI declares FP_* (fp_cli/app.py), and an unknown variable is not an error. Setting AGENTEYE_DASHBOARD_URL does not retarget the CLI; it is ignored and the command silently runs against the saved dashboard instead.AGENTEYE_HOME and AGENTEYE_ENVIRONMENT do still exist, but they belong to the collector and the telemetry SDK, not to this CLI.
Commands that delete, revoke, suppress, resolve, or replace configuration prompt by default. Use --yes only after verifying the active organization and target.