litellm_admin_tools
The connector’s action selects a read operation for the saved connection. For
credential setup, least-privilege guidance, and the distinction from model
routing, see
Connections.
Actions
Parameters
The new
anomaly_settings thresholds have these types and defaults:
GET_DAILY_KEY_SPEND returns one row per date, key hash, and model, with spend,
successful and failed request counts, and token counts; one key can therefore have
several rows per day. Its response also includes total_requests and
total_failed_requests summed across the returned rows. It reads one day at a time
from LiteLLM’s aggregated daily-activity route when available, and falls back to the
paged route when that route is unsupported or lacks per-model key detail.
Without dates, spend actions cover the last seven days, including today.
GET_SPEND_LOGS returns request rows without prompt or response text; older
proxies may return per-key totals instead. GET_DAILY_KEY_SPEND reads at most 20
pages per day on its paged fallback; when its result is truncated: true, narrow
the date range before using the partial totals. Credentials in model and
config results are redacted. For charts and recurring reports, use the
litellm_admin dataset source
instead of scheduling repeated tool calls.
DETECT_SPEND_ANOMALIES reads the last complete UTC day by default; set end_date to
choose another report end date. Its anomaly_settings.time_budget_seconds is a best-effort budget
(default 150, allowed 30–300): no read starts after the budget expires, and each
request timeout is cut to the remaining time. A read already in flight may finish later if
the proxy continues sending data. If the key list or daily spend cannot be read, the action
returns an error; failed optional reads are skipped and named in the report notes. A GET
that times out is retried once. The paged spend-log route is used when available, with a
legacy route fallback.
The rotated_key finding flags an unknown key hash whose alias now belongs to a listed
key. It is MEDIUM when the old key still succeeds or spends more than one hour after
the new key was created, or continues failing beyond
anomaly_settings.rotated_failures_medium_after_hours (default 48); otherwise it is LOW. If the new
key’s creation time is unavailable, successful requests or spend in the report window
prevent a LOW severity.
Unlisted hashes with spend or successful requests are checked with GET /key/info
(at most five lookups, highest spend first). A hash that exists but is absent from
the key list is reported as key_missing_from_list; failed lookups are noted and
do not confirm deletion. Traffic associated with a Genesis auto-provisioned source
is reported as auto_provisioned_key_gone: deletion-like evidence can be LOW,
while successful-only traffic, uncertain deletion timing, daily spend above
auto_provisioned_gone_medium_spend, traffic continuing past
auto_provisioned_gone_medium_after_hours, or a failed lookup is MEDIUM.
When the lookup is skipped, severity follows the available traffic evidence and
thresholds.
Unknown hashes with only failed requests and $0 spend are grouped into one
unknown_key_failed_only finding rather than one finding per hash. It lists the
available routes, user agents and requester IPs; it is LOW when every log read
is model discovery and MEDIUM for other calls or a sufficiently large daily
jump above the trailing median. A spend spike against a baseline below $0.01 is
described as new spend against an approximately-zero baseline, without a ratio.
For example, list keys and then inspect daily spend for one key:
key_hash as key with GET_DAILY_KEY_SPEND or
GET_KEY_INFO.
