Stats API

The Stats API exposes aggregated request metrics. Use GET /stats for the dashboard summary view (today, yesterday, last 7 days, last hour, last minute) and GET /stats/timeseries for time-bucketed data suitable for charts.

Base URL: https://<your-gateway-host>/admin/v1

GET /stats

Returns a snapshot of request counts, token usage, cost, and latency across several time windows. Also includes a per-tenant breakdown and the most recent log entries.

Every caller below admin/tenant_admin (member, viewer, ki_manager, demouser) is scoped to their own data. Non-admin callers are pinned to their own tenant server-side (a spoofed tenant_id is ignored); a caller with no tenant is refused 403 { "error": "forbidden" } (the query never runs). A platform admin may target one tenant with ?tenant_id= or omit it (and an empty value) for the global, all-tenant view.

The calendar-day windows (today, yesterday, last_7d) are computed in the viewer's timezone, supplied by the tz_offset query parameter — so "today" means the viewer's local calendar day, not a UTC day. The rolling windows (hour, last_min) are relative to the current instant and are timezone-independent.

Query parameters

Parameter	Type	Default	Description
`tenant_id`	string (UUID)	caller's tenant	Platform `admin` only: target one tenant, or omit / empty for the global view. Ignored (server-pinned) for every other role.
`tz_offset`	integer	`0` (UTC)	The viewer's UTC offset in minutes, in the JavaScript `Date.getTimezoneOffset()` convention: UTC − local, so Central European Summer Time (UTC+2) is `-120` and US Eastern Standard Time (UTC−5) is `300`. Used only to anchor the `today` / `yesterday` / `last_7d` calendar-day boundaries to the viewer's local midnight. Accepted: an integer within `[-840, 720]` (UTC−14 … UTC+14). Rejected → falls back to UTC (offset 0): any non-numeric, empty, repeated, out-of-range, or missing value (fractional values are floored toward −∞). It never causes an error and never influences authorization — it shifts a display window only.

The GET /stats/timeseries series that back the dashboard sparkline glyphs accept the same tz_offset parameter and bucket in the viewer's local frame — so the daily (1d) bucket and, for a fractional-hour offset (e.g. India, UTC+5:30), the hourly buckets align with the viewer's local day/hour rather than UTC. For whole-hour offsets (e.g. Central Europe, US) the hourly buckets coincide with local hours either way. Omit tz_offset (or send a rejected value) and buckets fall back to UTC.

curl "https://<your-gateway-host>/admin/v1/stats?tz_offset=-120"

Response structure

{
  "today":     { ...PeriodStats },
  "yesterday": { ...PeriodStats },
  "last_7d":   { ...PeriodStats },
  "hour":      { ...PeriodStats },
  "last_min":  { ...PeriodStats },
  "by_tenant": [ ...TenantStats ],
  "recent":    [ ...RecentEntry ],
  "recent_blocked": [ ...RecentBlockedEntry ]
}

recent and recent_blocked resemble LogEntry but are not the same shape. recent (up to 10 newest entries) omits id and instead adds the tenant and gateway slugs alongside the IDs. recent_blocked (up to 20 newest blocked/scrubbed/flagged entries) is a narrow subset carrying only the block/guardrail fields plus prompt_scrubbed and response_raw.

PeriodStats fields

Field	Type	Description
`requests`	integer	Total inference requests in the period.
`cached`	integer	Requests served from cache (no provider call made).
`blocked`	integer	Requests blocked by auth, rate limit, quota, or detectors.
`rate_limited`	integer	Requests that hit a rate limit in the period.
`scrubbed`	integer	Requests where a guardrail scrubbed PII (personally identifiable information) from the payload but allowed the request through.
`flagged`	integer	Requests where a guardrail raised a flag but took no blocking or scrubbing action.
`input_tokens`	integer	Total prompt tokens consumed.
`output_tokens`	integer	Total completion tokens generated.
`cost_usd`	number	Total cost in USD (from model pricing table).
`saved_cost_usd`	number	Cost saved by cache hits (would-have-been cost of cached requests).
`avg_latency_ms`	number	Average end-to-end request latency in milliseconds.
`avg_upstream_latency_ms`	number	Average time waiting for the upstream provider, excluding gateway overhead.

TenantStats fields (GET /stats)

by_tenant in GET /stats is a summary view with the following fields:

Field	Type	Description
`tenant_id`	string	Tenant UUID.
`tenant`	string	Tenant slug.
`requests`	integer	Total requests.
`input_tokens`	integer	Total input tokens.
`output_tokens`	integer	Total output tokens.
`cost_usd`	number	Total cost in USD.
`budget_remaining`	number	Remaining budget in USD for the tenant's current period (the live-monitor "Quota Left" column). Absent when the tenant has no budget.
`budget_remaining_unknown`	boolean	`true` when the remaining budget could not be computed (a transient spend-read failure) — distinct from "no limit".

The full TenantStats shape (with blocked, cached, errors, avg_latency_ms, etc.) is returned by GET /stats/analytics. See below.

Example response

{
  "today": {
    "requests": 1423,
    "cached": 82,
    "blocked": 14,
    "input_tokens": 1840200,
    "output_tokens": 312400,
    "cost_usd": 9.84,
    "saved_cost_usd": 0.41,
    "avg_latency_ms": 820,
    "avg_upstream_latency_ms": 760
  },
  "yesterday": {
    "requests": 2101,
    "cached": 134,
    "blocked": 22,
    "input_tokens": 2750000,
    "output_tokens": 480000,
    "cost_usd": 14.30,
    "saved_cost_usd": 0.67,
    "avg_latency_ms": 905,
    "avg_upstream_latency_ms": 850
  },
  "last_7d": { "requests": 11200, "cost_usd": 68.12, "..." : "..." },
  "hour":    { "requests": 61, "cost_usd": 0.38, "...": "..." },
  "last_min": { "requests": 2, "cost_usd": 0.01, "...": "..." },
  "by_tenant": [
    {
      "tenant_id": "ten_abc123",
      "tenant": "myapp",
      "requests": 1423,
      "input_tokens": 1840200,
      "output_tokens": 312400,
      "cost_usd": 9.84
    }
  ],

  "recent": [ { "...": "LogEntry" } ],
  "recent_blocked": [ { "...": "LogEntry" } ]
}

For the related LogEntry field definitions see the Logs API. Note that the recent and recent_blocked rows are not full LogEntry objects (see the field notes above).

GET /stats/analytics

Returns latency percentiles, top-model breakdown, and per-tenant, per-gateway, per-channel, per-agent, and per-user cost summaries, plus activity counters. Used by the analytics dashboard view.

On the global, all-tenant view (a platform admin who omits tenant_id), every grouping and counter excludes synthetic fixture tenants — e.g. the e2e-synthetic monitoring prober — so the platform analytics reflect real usage. A tenant-scoped view (?tenant_id=…) still returns them.

curl "https://<your-gateway-host>/admin/v1/stats/analytics?since=1742544000000"

Query parameters

Parameter	Type	Default	Description
`since`	integer	24 hours ago	Start of the analysis window as Unix milliseconds.
`until`	integer	now	Optional upper bound as Unix milliseconds. Used for closed windows such as "yesterday".
`tenant_id`	string	—	Filter by tenant. Honoured only for callers with the `admin` role. Non-admin callers are always scoped to their own tenant.
`anonymize`	boolean	`true` (anonymized)	Anonymized by default. Unless the caller explicitly opts out with `anonymize=0` (or `false`), the response omits the per-user grouping (`by_user`) entirely — no individual identifiers (email/user_id) are returned. Any other value (absent, `1`, `true`, or malformed) keeps the response anonymized (fail-closed). Aggregate groupings (channel, agent, model, gateway, tenant) and counters are always included. The flag is scoped to this endpoint. Note: in a single-user tenant, per-agent/per-model aggregates can still imply one person's usage.
`group_id`	string	—	Restrict the whole response to the members of one user group. For a KI-Manager this must be a group they manage (see below); `admin`/`tenant_admin` may narrow to any group in their tenant scope. An unknown or out-of-scope `group_id` (including a `group_id` paired with a mismatched `tenant_id`) returns `404 not found`; a KI-Manager requesting a same-tenant group they do not manage returns `403 forbidden`.
`tz_offset`	integer	`0` (UTC)	The viewer's UTC offset in minutes, in the JavaScript `Date.getTimezoneOffset()` convention: UTC − local (Central European Summer Time = `-120`, US Eastern Standard Time = `300`). Applies only to the per-agent daily series (`by_agent_series`), whose buckets are floored in the viewer's local frame so each daily bar starts at the viewer's local midnight — aligning it with the calendar-day window (`since`/`until`) the caller computes with the same offset. Accepted: an integer within `[-840, 720]` (UTC−14 … UTC+14). Rejected → falls back to UTC (offset 0): any non-numeric, empty, repeated, out-of-range, or missing value (fractional values are floored toward −∞). It never causes an error and never influences authorization — it shifts the bucket boundaries only.

💡 Note: /stats/analytics requires the admin, tenant_admin, or ki_manager role; lower-privileged callers (member, viewer, demouser) receive 403 forbidden. The anonymize=0 opt-out is therefore confined to those privileged roles, and only ever reveals users within the caller's authorized scope: admin/tenant_admin see their own tenant (a tenant_admin is pinned to it), while a ki_manager is pinned to the members of the groups they manage. Any role that could send anonymize=0 without authorization is already blocked with 403 before the handler runs; the client value never decides access.

KI-Manager group scoping

A KI-Manager may optionally see the statistics of the groups they manage. Unlike admin/tenant_admin (who see the whole tenant/workspace), a KI-Manager is always scoped to a group's members and can never see tenant-wide or another group's data:

"Groups they manage" are the groups the KI-Manager created (user_group.created_by). This is the same boundary as the group-management API (can_manage_group). Limitation: the data model has no separate "group manager" or delegation concept, so a group created by an admin and handed to a KI-Manager to run — or a group whose creator account was later deleted (created_by becomes NULL) — is not visible to the KI-Manager. Only admin/tenant_admin see those.
Without group_id: the response covers the union of the members of all groups the KI-Manager manages. A KI-Manager who manages no (non-empty) groups gets an all-zero response — never a tenant-wide fallback.
With group_id: the response covers only that group's members, and only if the KI-Manager manages it (else 403/404 as above).
"A group's statistics" = activity by the group's members. All groupings and counters are filtered to request_log rows / rows owned by the member set. Because request_log carries no group dimension, this is the only coherent definition. Consequences to keep in mind:
by_agent is filtered on the run-as owner (user_id) — which, because invoke is owner-only, is the real member who invoked. Rows roll up over the clone lineage (see AgentStats), so a KI-Manager sees, for each shared agent, the invocations/conversations/distinct-users contributed by their members' clones of it (the display name is the root agent's, which may be owned outside the group).
active_users is windowed: under a group filter it is the count of the group's members active in the window (distinct members with ≥1 request in the period), not the group's current member count.

Response structure

{
  "percentiles": { ...LatencyPercentiles },
  "top_models":  [ ...TopModelRow ],
  "by_tenant":   [ ...TenantStats ],
  "by_gateway":  [ ...GatewayStats ],
  "by_channel":  [ ...ChannelStats ],
  "by_agent":    [ ...AgentStats ],
  "by_agent_series": [ ...AgentSeries ],
  "by_user":     [ ...UserStats ],
  "counters":    { ...ActivityCounters },
  "anonymized":  true,
  "cache_efficiency": { ...CacheEfficiency }
}

ChannelStats fields (UI vs API vs agent)

Each row is the usage attributed to one access channel. The channel is derived: agent for agent invocations, ui for chat-session (playground) tokens, api otherwise.

Field	Type	Description
`channel`	string	`ui`, `api`, or `agent`.
`requests`	integer	Request count for the channel.
`blocked`	integer	Requests blocked or scrubbed by a guardrail.
`cached`	integer	Requests served from the response cache.
`input_tokens`	integer	Total prompt tokens consumed on the channel.
`output_tokens`	integer	Total completion tokens produced on the channel.
`cost_usd`	number	Total cost.
`saved_cost_usd`	number	Cost avoided by cache hits.
`errors`	integer	HTTP ≥ 400 (excluding client-cancel 499).
`client_aborted`	integer	Requests the client cancelled (HTTP 499).
`avg_latency_ms`	integer	Average latency for non-blocked requests.

AgentStats fields (per agent, rolled up over the clone lineage)

Each row is one lineage root — a shared/original agent plus every clone made from it. Rows are keyed by agent_id = the root's id; a clone's usage rolls into its root (the counting model below). Filtering (tenant, KI-Manager group scope) applies before the roll-up.

Field	Type	Description
`agent_id`	string	The lineage root agent's id (a shared agent + its clones share one row).
`agent`	string	Root agent name (or slug). A hard-deleted root falls back to the raw id.
`invocations`	integer	Invocations across the lineage in the window.
`conversations`	integer	Distinct chat sessions the lineage was used in. Headless invokes (API/schedule) carry no session id and count as invocations, not conversations — so this is 0 for agents only ever invoked headlessly, and becomes non-zero when an agent is used inside a chat thread. Read from `request_log_legs.conversation_id`; no per-invoke session id is fabricated.
`distinct_users`	integer	Distinct people who used the lineage — see the counting model below.
`cost_usd`	number	Total cost across the lineage.

Per-agent "distinct users" — the counting model (supersedes the earlier model). Invoke is owner-only (agent_invoke refuses a non-owner caller), and the catalog "use" action is clone-then-own (a member who wants a shared agent clones it into a new agent they own, then invokes that). So a single agent's distinct-invoker count is structurally ≤1 — the "Scheinwert". The fix is not to drop the metric but to count across the clone lineage: each member who uses a shared agent owns a distinct clone and invokes it as owner (invoker_id = that member), so COUNT(DISTINCT invoker_id) over the root and its clones = the true number of people using it. A clone records its lineage root in agent.source_agent_id (set by the clone flow, flattened so the tree is one level); every request stamps request_log.agent_root_id, so the count survives the clone being deleted. Caveats (bid-relevant): lineage is recorded from the release that introduced it — pre-existing clones are not retro-linked; and because there is no server clone endpoint yet, the lineage pointer is client-declared within the caller's visibility (a member can omit it → undercount, or attribute a clone to any agent they can already see) — so distinct_users is a best-effort, self-reported figure, not server-authoritative. Unattended runs (scheduler/service token) log a NULL invoker and are excluded from the count.

AgentSeries fields (per-agent calls over time)

One entry per top agent (up to 8, by invocation count) carrying a daily-bucketed invocation time series. The series is dense and zero-filled across the window, so every day is present even with no calls. Timestamps are milliseconds since epoch (bucket start); when the tz_offset query parameter is supplied the bucket start is the viewer's local midnight rather than UTC midnight (so the bars align with the viewer-local calendar-day windows). Empty ([]) when there are no agent invocations in the window.

Field	Type	Description
`agent_id`	string	The saved agent's id.
`agent`	string	Agent name (or slug).
`series`	array	Daily points: `{ "ts": <ms>, "invocations": <integer> }`.

ActivityCounters fields

Counts for the window, scoped to the tenant filter. *_created count items created in the window; active_users counts distinct users active in the window — users with at least one logged request (UI or API channel) between since and until. System/unattributed requests (no user_id) are excluded; duplicate requests by one user count once. It is not the provisioned-user headcount, so it can differ from the number of user rows (e.g. a user active in the window then deleted still counts).

Two caveats follow from active_users being derived from request_log: - No explicit window (as in the tenant export's usage-statistics snapshot) → active_users reflects the trailing 24h, consistent with the sibling *_created counters, which use the same default. The provisioned headcount remains available separately (e.g. the export's users array). - Request logging disabled for a tenant (request_logging_disabled) → no request_log rows are written, so active_users reads 0 even while conversations/messages (sourced from the chat tables) show activity.

Field	Type
`conversations`	integer
`messages`	integer
`agents_created`	integer
`projects_created`	integer
`documents_created`	integer
`active_users`	integer (distinct users active in the window)

LatencyPercentiles fields

Field	Type	Description
`p50`	number \| null	Median end-to-end latency in milliseconds. `null` if no data.
`p95`	number \| null	95th-percentile latency in milliseconds.
`p99`	number \| null	99th-percentile latency in milliseconds.

Percentiles cover only non-blocked requests.

TopModelRow fields

Field	Type	Description
`model`	string	Model name.
`provider`	string	Provider slug (e.g. `openai`, `anthropic`).
`requests`	integer	Request count in the window.
`blocked`	integer	Requests blocked or scrubbed by a guardrail.
`cached`	integer	Requests served from the response cache.
`input_tokens`	integer	Total prompt tokens consumed by the model.
`output_tokens`	integer	Total completion tokens produced by the model.
`cost_usd`	number	Total cost in USD.
`saved_cost_usd`	number	Cost avoided by cache hits.
`errors`	integer	HTTP ≥ 400 (excluding client-cancel 499).
`client_aborted`	integer	Requests the client cancelled (HTTP 499).
`avg_latency_ms`	number \| null	Average end-to-end latency (ms) over non-blocked requests; `null` if every request was blocked.

Up to 10 models are returned, ordered by request count descending.

GatewayStats fields

Field	Type	Description
`gateway_id`	string	Gateway UUID.
`gateway`	string	Gateway slug.
`tenant`	string \| null	Tenant slug.
`requests`	integer	Total requests.
`blocked`	integer	Requests blocked or scrubbed (counts `blocked=1` OR `scrub_applied=1`).
`cached`	integer	Cache hit count.
`input_tokens`	integer	Total input tokens.
`output_tokens`	integer	Total output tokens.
`cost_usd`	number	Total cost in USD.
`saved_cost_usd`	number	Cost saved by cache hits.
`avg_latency_ms`	number	Average end-to-end latency in milliseconds.
`errors`	integer	Requests that returned an HTTP status `>= 400`, excluding `499` client-aborted.
`client_aborted`	integer	Requests the client aborted before response headers (HTTP `499`). A stream cancelled after headers (Stop pressed mid-answer) is a `status` `200` row with `aborted` `1` in the request log and is not counted here.

UserStats fields

Field	Type	Description
`user_id`	string	User UUID (from the `user_id` field of the auth token).
`tenant_id`	string	Tenant UUID the user's requests belong to.
`email`	string \| null	User email, if known.
`requests`	integer	Total requests attributed to this user.
`blocked`	integer	Requests blocked or scrubbed (counts `blocked=1` OR `scrub_applied=1`).
`cached`	integer	Cache hit count.
`input_tokens`	integer	Total input tokens.
`output_tokens`	integer	Total output tokens.
`cost_usd`	number	Total cost in USD.
`saved_cost_usd`	number	Cost saved by cache hits.
`avg_latency_ms`	number	Average end-to-end latency in milliseconds.
`errors`	integer	Requests that returned an HTTP status `>= 400`, excluding `499` client-aborted.
`client_aborted`	integer	Requests the client aborted before response headers (HTTP `499`). A stream cancelled after headers (Stop pressed mid-answer) is a `status` `200` row with `aborted` `1` in the request log and is not counted here.

by_user is empty by default (anonymized) and is populated only when the caller explicitly opts out with anonymize=0. When populated, it only includes requests made with auth tokens that have a user_id set; anonymous token requests are not included. Up to 50 users are returned, ordered by cost descending. The example below shows the opted-out (identifying) response — the default response has "by_user": [] and "anonymized": true.

CacheEfficiency fields

cache_efficiency summarises input-token caching and input cost over the window (successful requests across all providers). Providers with prompt caching (Anthropic) populate the cache-read/write terms; providers without it (e.g. local models) contribute only their standard input cost at a 0 % hit rate.

Field	Type	Description
`cache_write_tokens`	integer	Prompt-cache write tokens — the total across both TTL tiers (the 1-hour part is counted once; it is a share of the total, not an addition to it).
`cache_read_tokens`	integer	Prompt-cache read tokens.
`standard_input_tokens`	integer	Non-cached input (prompt) tokens.
`uncached_cost_usd`	number	Cost that would have applied without the prompt cache.
`cached_cost_usd`	number	Cost actually incurred reading from the cache.
`cache_hit_pct`	number \| null	Share of input tokens served from cache, as a percentage. `null` when there is no data.
`compaction_tokens_saved`	integer	Input tokens saved by conversation compaction.
`compaction_cost_saved`	number	Cost saved by conversation compaction in USD.

Example response

{
  "percentiles": { "p50": 420, "p95": 1840, "p99": 3210 },
  "top_models": [
    { "model": "gpt-4o", "provider": "openai", "requests": 842,
      "input_tokens": 1120400, "output_tokens": 208900, "cost_usd": 3.14, "avg_latency_ms": 680 }
  ],
  "by_tenant": [
    {
      "tenant_id": "ten_abc123", "tenant": "myapp",
      "requests": 1423, "blocked": 14, "cached": 82,
      "input_tokens": 1840200, "output_tokens": 312400,
      "cost_usd": 9.84, "saved_cost_usd": 0.41,
      "avg_latency_ms": 820, "errors": 3, "client_aborted": 1
    }
  ],
  "by_gateway": [
    {
      "gateway_id": "gw_xyz789", "gateway": "prod", "tenant": "myapp",
      "requests": 1423, "blocked": 14, "cached": 82,
      "input_tokens": 1840200, "output_tokens": 312400,
      "cost_usd": 9.84, "saved_cost_usd": 0.41,
      "avg_latency_ms": 820, "errors": 3, "client_aborted": 1
    }
  ],
  "by_channel": [
    { "channel": "ui",    "requests": 900, "input_tokens": 1200000, "output_tokens": 210000,
      "cost_usd": 6.10, "errors": 2, "avg_latency_ms": 610 },
    { "channel": "api",   "requests": 400, "input_tokens": 560000,  "output_tokens": 92000,
      "cost_usd": 3.20, "errors": 1, "avg_latency_ms": 700 },
    { "channel": "agent", "requests": 123, "input_tokens": 80200,   "output_tokens": 10400,
      "cost_usd": 0.54, "errors": 0, "avg_latency_ms": 540 }
  ],
  "by_agent": [
    { "agent_id": "agt_support", "agent": "Support Bot", "invocations": 90,
      "conversations": 12, "distinct_users": 4, "cost_usd": 0.54 }
  ],
  "by_agent_series": [
    { "agent_id": "agt_support", "agent": "Support Bot", "series": [
      { "ts": 1742457600000, "invocations": 12 },
      { "ts": 1742544000000, "invocations": 0 },
      { "ts": 1742630400000, "invocations": 31 }
    ] }
  ],
  "by_user": [
    {
      "user_id": "usr_alice", "tenant_id": "ten_abc123", "email": "alice@myapp.example",
      "requests": 341, "blocked": 2, "cached": 18,
      "input_tokens": 440200, "output_tokens": 74800,
      "cost_usd": 2.31, "saved_cost_usd": 0.09,
      "avg_latency_ms": 790, "errors": 1, "client_aborted": 0
    }
  ],
  "anonymized": false,
  "cache_efficiency": {
    "cache_write_tokens": 12800, "cache_read_tokens": 90400,
    "standard_input_tokens": 1737000,
    "uncached_cost_usd": 27.40, "cached_cost_usd": 0.27,
    "cache_hit_pct": 4.9,
    "compaction_tokens_saved": 0, "compaction_cost_saved": 0
  }
}

GET /stats/timeseries

Returns an array of time-bucketed data points, suitable for rendering sparklines or charts. Buckets with no activity are included as zero-filled entries so the array length is always exactly n.

Scope follows GET /stats: non-admin callers are pinned to their own tenant and own data server-side; a caller with no tenant is refused 403. A platform admin may pass ?tenant_id= to target one tenant, or omit it (empty value included) for the global view. The global series excludes synthetic fixture tenants (e.g. the e2e-synthetic monitoring prober); a tenant-scoped series includes them.

curl "https://<your-gateway-host>/admin/v1/stats/timeseries?bucket=1h&n=24"

Query parameters

Parameter	Type	Default	Description
`bucket`	string	`1h`	Bucket size. One of: `5m`, `15m`, `30m`, `1h`, `6h`, `1d`.
`n`	integer	`24`	Number of buckets to return. Range: 1–168.
`until`	integer	now	End of the time range as a Unix timestamp in seconds. Defaults to the current time.
`tz_offset`	integer	`0` (UTC)	The viewer's UTC offset in minutes, in the JavaScript `Date.getTimezoneOffset()` convention: UTC − local (Central European Summer Time = `-120`, US Eastern Standard Time = `300`). Buckets are floored in the viewer's local frame, so the `1d` bucket starts at local midnight and a fractional-hour offset (e.g. UTC+5:30) shifts the hourly buckets to local boundaries. Accepted: an integer within `[-840, 720]` (UTC−14 … UTC+14). Rejected → falls back to UTC (offset 0): any non-numeric, empty, repeated, out-of-range, or missing value (fractional values are floored toward −∞). It never causes an error and never influences authorization — it shifts the bucket boundaries only.

TimeseriesPoint fields

Field	Type	Description
`ts`	integer	Bucket start time as Unix milliseconds.
`requests`	integer	Total requests in the bucket.
`blocked`	integer	Requests blocked in the bucket.
`rate_limited`	integer	Requests rate-limited in the bucket.
`errors`	integer	Requests that returned an HTTP status `>= 400` (excluding `499` client-aborted) in the bucket.
`client_aborted`	integer	Requests the client aborted before response headers (HTTP `499`) in the bucket. Streams cancelled after headers are `status` `200` rows with `aborted` `1` in the request log, not counted here.
`cost_usd`	number	Total cost in USD for the bucket.

💡 Note: ts is in Unix milliseconds (not seconds) for direct compatibility with charting libraries that expect millisecond-precision timestamps (e.g. Chart.js, Recharts, Grafana).

Example response

[
  {"ts": 1742540400000, "requests": 42, "blocked": 1, "rate_limited": 0, "errors": 1, "client_aborted": 0, "cost_usd": 0.0314},
  {"ts": 1742544000000, "requests": 67, "blocked": 0, "rate_limited": 0, "errors": 0, "client_aborted": 0, "cost_usd": 0.0521},
  {"ts": 1742547600000, "requests": 0,  "blocked": 0, "rate_limited": 0, "errors": 0, "client_aborted": 0, "cost_usd": 0.0},
  {"ts": 1742551200000, "requests": 18, "blocked": 2, "rate_limited": 1, "errors": 2, "client_aborted": 1, "cost_usd": 0.0119}
]

GET /health/dashboard

GET /admin/v1/health/dashboard

Returns the aggregated operational-health summary that the Health dashboard renders in a single call — the synthetic-probe results, the last-24-hours non-2xx and client-aborted request counts, the phoned-home client errors, and the other failure surfaces shown there.

Requires the HEALTH_VIEW permission; the Health dashboard itself is available to the admin role only. A backend failure returns 500.

Platform web-search key (AGF-2874). The summary carries web_search_platform_key_missing (boolean): true when the platform Linkup key (the trial_linkup_api_key setting — applied to every keyless Linkup gateway on every plan) is unset while at least one production gateway of a non-deleted tenant carries a keyless Linkup web_search block — i.e. web search is silently dark for it. false when the key is set, or when nothing depends on it yet. The companion object web_search_platform_key gives the inputs: { "key_set": boolean, "keyless_production_gateways": integer }. If the count could not be read, all three are absent and web_search_platform_key_error (string) names the fault — a failed read is never reported as healthy; the dashboard renders "unknown". The same condition drives the [platform_alert] signal web_search_platform_key_missing (checked every 10 minutes, ERR-logged each check, notified at most every 6 hours).

Plan-aware model copy (AGF-2930). The summary carries plan_model_copy_stale (boolean): true when, for at least one offered model set served in the last 7 days, the set's current copy generation was rejected by the validator (or the set's catalog facts could not be read as valid input) — the model picker then shows that set's hand-written seed line, a checked catalog tagline or the model name instead (see Tagline resolution). plan_model_copy_indistinct (boolean) is true when a current set offers two models no fact tells apart (same price rank, context, capabilities, hosting and speed) — a plan-design problem, not a generator fault. The companion object plan_model_copy gives the counts: { "stale_sets": integer, "indistinct_sets": integer }. If the status could not be read, all three are absent and plan_model_copy_error (string) names the fault — the dashboard renders "unknown", never healthy.

Cost-of-goods (COGS)

The COGS endpoints aggregate the per-leg billing ledger (request_log_legs) into cost breakdowns by model, by tenant, by service, or by API key. A single inference request may produce several legs (the primary call, any fallback, plus side services such as guardrails or web search), so these views give a finer-grained cost picture than the request-level Stats summary.

All four endpoints share the same time-window selection:

Query parameters (all COGS endpoints)

Parameter	Type	Default	Description
`range`	string	`30d`	Relative window ending now. Format `<n><unit>` where unit is `d` (days), `h` (hours), or `w` (weeks) — e.g. `24h`, `2w`, `90d`. Ignored when `from`/`to` are given.
`from`	integer	—	Explicit window start as a Unix timestamp in seconds. Must be paired with `to`.
`to`	integer	—	Explicit window end as a Unix timestamp in seconds. Must be paired with `from`.

💡 Note: When neither from/to nor range is given, the window defaults to the last 30 days plus one day into the future (the extra day captures in-flight legs still being written).

CSV export (`format`) — cost-allocation endpoints

The three cost-allocation endpoints — /cogs/per-token, /cogs/per-user, and /cogs/per-agent — accept a format parameter for a machine-readable download:

Parameter	Type	Default	Description
`format`	string	`json`	Response encoding. `json` (or absent) returns the JSON array documented per endpoint. `csv` returns the same rows as `text/csv; charset=utf-8` with a fixed header row (the response-field order below), served as a file attachment (`Content-Disposition: attachment`). Any other value — an unknown string, a repeated `?format=` (which arrives as a list), or a bare `?format` (no value) — is rejected with `400 { "error": "invalid format (json or csv)" }` (fail-closed allowlist), validated before the aggregate query runs.

Notes on the CSV encoding:

The column order is fixed and equals the endpoint's response-field order; a missing value renders as an empty cell.
Cells are RFC-4180 quoted, and any value that begins with =, +, -, @, TAB, or CR is prefixed with a single quote to neutralize spreadsheet formula injection (CWE-1236) — so a user email or agent name such as =SUM(...) cannot execute when the file is opened in a spreadsheet.
No byte-order mark (BOM) is emitted.
The download filename is a fixed per-endpoint constant (cogs-per-token.csv / cogs-per-user.csv / cogs-per-agent.csv); the server sets Access-Control-Expose-Headers: Content-Disposition so a browser SPA fetching cross-origin can read it.
The bucket, tenant_id, scope, and billable_only rules are identical to the JSON response; format only changes the encoding.

GET /cogs/per-model

Cost and token totals grouped by (provider, model, tier, byok), ordered by cost descending. Required role: admin or tenant_admin.

curl "https://<your-gateway-host>/admin/v1/cogs/per-model?range=7d"

Additional query parameters

Parameter	Type	Default	Description
`tenant_id`	string	—	Restrict to one tenant. Honoured only for callers with the `admin` role; every other caller (including `tenant_admin`) is scoped to their own tenant server-side, and a caller with no tenant is refused `403`.
`billable_only`	string	`1`	Include only billable legs. Set to `0` to also include non-billable legs (e.g. free-tier or internal).

Response

Array of rows. Returns 500 { "error": ... } on a storage failure.

Field	Type	Description
`provider`	string	Provider that served the leg.
`model`	string	Model used for the leg.
`tier`	string \| null	Pricing tier recorded for the leg.
`byok`	integer	`1` if the leg used a bring-your-own-key credential, `0` otherwise.
`call_count`	integer	Number of legs in this group.
`in_t`	integer	Sum of input (prompt) tokens.
`out_t`	integer	Sum of output (completion) tokens.
`cache_read_t`	integer	Sum of prompt-cache read tokens.
`cache_write_t`	integer	Sum of prompt-cache creation (write) tokens.
`cost_usd`	number	Total cost in USD for the group.
`partial_count`	integer	Number of legs flagged partial (interrupted / incomplete).

[
  {
    "provider": "anthropic",
    "model": "claude-opus-4-6",
    "tier": null,
    "byok": 0,
    "call_count": 412,
    "in_t": 1840200,
    "out_t": 312400,
    "cache_read_t": 90400,
    "cache_write_t": 12800,
    "cost_usd": 9.84,
    "partial_count": 2
  }
]

GET /cogs/per-tenant

Cost and token totals grouped by tenant, billable legs only, ordered by cost descending. Required role: platform admin only — lower-privileged callers receive 403 { "error": "forbidden: platform admin only" }.

curl "https://<your-gateway-host>/admin/v1/cogs/per-tenant?range=30d"

Response

Array of rows. Returns 500 { "error": ... } on a storage failure.

Field	Type	Description
`tenant_id`	string	Tenant UUID.
`call_count`	integer	Number of billable legs for the tenant.
`cost_usd`	number	Total cost in USD.
`in_t`	integer	Sum of input tokens.
`out_t`	integer	Sum of output tokens.

[
  { "tenant_id": "ten_abc123", "call_count": 1423, "cost_usd": 14.30, "in_t": 2750000, "out_t": 480000 }
]

GET /cogs/per-service

Per-service usage grouped by (provider, service, unit_kind), ordered by call count descending. Counts only legs that carry a non-null service (e.g. web_search, guardrail). Required role: admin or tenant_admin.

curl "https://<your-gateway-host>/admin/v1/cogs/per-service?range=30d"

Additional query parameters

Parameter	Type	Default	Description
`tenant_id`	string	—	Restrict to one tenant. Honoured only for callers with the `admin` role. Every other caller (including `tenant_admin`) is scoped to their own tenant server-side; a spoofed `tenant_id` is ignored, and a caller with no tenant is refused `403`. A caller can never read another tenant's — or the global, all-tenant — per-service totals.

Plus the shared range / from / to window parameters above.

Response

Array of rows. Returns 500 { "error": ... } on a storage failure.

Field	Type	Description
`provider`	string	Provider that served the service leg.
`service`	string	Service name (e.g. `web_search`, `guardrail`).
`unit_kind`	string \| null	Unit the service is metered in (e.g. `requests`, `tokens`).
`call_count`	integer	Number of legs in this group.
`units_total`	number	Sum of metered units.
`avg_latency_ms`	number	Average leg latency in milliseconds.
`in_t`	integer	Sum of input tokens (where applicable).
`out_t`	integer	Sum of output tokens (where applicable).

[
  { "provider": "anthropic", "service": "web_search", "unit_kind": "requests", "call_count": 88, "units_total": 88, "avg_latency_ms": 640, "in_t": 0, "out_t": 0 }
]

GET /cogs/per-token

Cost and usage aggregated per API key (token) × period (Zeitraum) × KI-Modell, ordered by cost descending. This is the billing-disclosure view: it answers "what did each API key spend, on which model, in which period." Required role: admin or tenant_admin.

curl "https://<your-gateway-host>/admin/v1/cogs/per-token?range=30d&bucket=month"

Only genuine LLM legs are counted: legs with a NULL model (non-LLM side services such as PII scrubbing or guardrails) are excluded, so the per-model figures are clean. Customer-paid (BYOK) legs are kept on their own rows via the byok dimension and are never summed into platform cost. A single user request that fans out into several same-model legs (e.g. a tool loop) counts as one request.

Additional query parameters

Parameter	Type	Default	Description
`bucket`	string	`month`	Period granularity for the `period` column. One of `day`, `week`, `month`, `total`. Any other value is rejected with `400 { "error": "invalid bucket" }` (fail-closed allowlist — the value is never interpolated into SQL). `total` collapses the whole window into a single period labelled `"total"`.
`tenant_id`	string	—	Restrict to one tenant. Honoured only for callers with the `admin` role. Every other caller (including `tenant_admin`) is scoped to their own tenant server-side; a spoofed `tenant_id` is ignored, and a caller with no tenant is refused `403`. A caller can never read another tenant's cost rows.
`token_id`	string	—	Restrict to a single API key. Must be a single scalar value — a repeated param (`?token_id=a&token_id=b`) is rejected with `400 { "error": "token_id must be a single value" }` (fail-closed; it is never bound into SQL). Combined with the mandatory tenant scope: a `token_id` belonging to another tenant returns an empty result.
`billable_only`	string	`1`	Include only billable legs. Set to `0` to also include non-billable legs (e.g. free-tier / self-hosted). Non-billable legs still obey the `model IS NOT NULL` rule.

Plus the shared range / from / to window parameters above.

Response

Array of rows. Returns 500 { "error": ... } on a storage failure, 400 { "error": "invalid bucket" } on a bad bucket.

Field	Type	Description
`token_id`	string	API-key id the legs were attributed to. Empty string (`""`) means non-API-key traffic — interactive UI / playground sessions that carry no API key.
`token_label`	string	Human label of the API key (from `auth_token.label`). Empty string if the key has no label or was deleted.
`period`	string	Period bucket. Format depends on `bucket`: `2026-07-15` (day), `2026-W29` (ISO year-week), `2026-07` (month), or `total`. Boundaries are computed in the gateway database session timezone (UTC on the standard deployment). Note: the `week` bucket uses the ISO week-numbering year, so late-December usage can fall in the following calendar year's `W01` (e.g. `2025-12-31` → `2026-W01`) — while `day`/`month` use the calendar year.
`provider`	string	Provider that served the model.
`model`	string	Model id (never null — non-model side-service legs are excluded).
`byok`	integer	`1` if the legs used a bring-your-own-key credential (customer-paid), `0` for platform-paid. Kept as a separate dimension so customer and platform cost never merge.
`requests`	integer	Distinct end-user requests in this group (`COUNT(DISTINCT request_log_id)`), not the leg count. This is a per-row count and is not additive across rows: a single request that fans out across models (a fallback) or across a period boundary is counted once in each row it touches, so summing the `requests` column overstates the true request total.
`input_tokens`	integer	Sum of input (prompt) tokens.
`output_tokens`	integer	Sum of output (completion) tokens.
`total_tokens`	integer	`input_tokens + output_tokens`.
`cost_usd`	number	Total cost in USD for the group, rounded to 6 decimals.

[
  {
    "token_id": "8f3c2a10-6b4e-4d21-9a77-1e5c0b9d4a02",
    "token_label": "prod-key",
    "period": "2026-07",
    "provider": "anthropic",
    "model": "claude-opus-4-7",
    "byok": 0,
    "requests": 412,
    "input_tokens": 1840200,
    "output_tokens": 312400,
    "total_tokens": 2152600,
    "cost_usd": 9.84
  }
]

GET /cogs/per-user

Cost and usage aggregated per user × period (Zeitraum) × KI-Modell, ordered by cost descending. This is the per-user cost-allocation view: it answers "which user spent how much, on which model, in which period." Required role: admin, tenant_admin, or finance (the COGS_PER_USER_VIEW permission). It is the exact cost twin of /cogs/per-token — same period buckets, same tenant scoping, same row shape — with the grouping key swapped from the API key to the user.

curl "https://<your-gateway-host>/admin/v1/cogs/per-user?range=30d&bucket=month"

Only genuine LLM legs are counted: legs with a NULL model (non-LLM side services such as PII scrubbing or guardrails) are excluded, so the per-model figures are clean. Customer-paid (BYOK) legs are kept on their own rows via the byok dimension and are never summed into platform cost. A single user request that fans out into several same-model legs (e.g. a tool loop) counts as one request.

Cost allocation is inherently per-identity: this endpoint returns each user's email within the caller's scope. A tenant_admin can only ever see the users of their own tenant (server-side tenant scope, fail-closed); only an admin may scope to another tenant or read across all tenants.

Additional query parameters

Parameter	Type	Default	Description
`bucket`	string	`month`	Period granularity for the `period` column. One of `day`, `week`, `month`, `total`. Any other value is rejected with `400 { "error": "invalid bucket" }` (fail-closed allowlist — the value is never interpolated into SQL). `total` collapses the whole window into a single period labelled `"total"`.
`tenant_id`	string	—	Restrict to one tenant. Honoured only for callers with the `admin` role. Every other caller (including `tenant_admin`) is scoped to their own tenant server-side; a spoofed `tenant_id` is ignored, and a caller with no tenant is refused `403`. A caller can never read another tenant's cost rows.
`user_id`	string	—	Restrict to a single user. Must be a single scalar value — a repeated param (`?user_id=a&user_id=b`) is rejected with `400 { "error": "user_id must be a single value" }` (fail-closed; never bound into SQL). Combined with the mandatory tenant scope: a `user_id` belonging to another tenant returns an empty result.
`billable_only`	string	`1`	Include only billable legs. Set to `0` to also include non-billable legs (e.g. free-tier / self-hosted). Non-billable legs still obey the `model IS NOT NULL` rule.

Plus the shared range / from / to window parameters above.

Response

Array of rows. Returns 500 { "error": ... } on a storage failure, 400 { "error": "invalid bucket" } on a bad bucket.

Field	Type	Description
`user_id`	string	User id the legs were attributed to. Empty string (`""`) means traffic with no attributed user — e.g. API-key requests that carry no user context.
`user_email`	string	Email of the user (from `user.email`). Empty string if the user has no email or was deleted.
`period`	string	Period bucket. Format depends on `bucket`: `2026-07-15` (day), `2026-W29` (ISO year-week), `2026-07` (month), or `total`. Boundaries are computed in the gateway database session timezone (UTC on the standard deployment). Note: the `week` bucket uses the ISO week-numbering year, so late-December usage can fall in the following calendar year's `W01` (e.g. `2025-12-31` → `2026-W01`) — while `day`/`month` use the calendar year.
`provider`	string	Provider that served the model.
`model`	string	Model id (never null — non-model side-service legs are excluded).
`byok`	integer	`1` if the legs used a bring-your-own-key credential (customer-paid), `0` for platform-paid. Kept as a separate dimension so customer and platform cost never merge.
`requests`	integer	Distinct end-user requests in this group (`COUNT(DISTINCT request_log_id)`), not the leg count. This is a per-row count and is not additive across rows: a single request that fans out across models (a fallback) or across a period boundary is counted once in each row it touches, so summing the `requests` column overstates the true request total.
`input_tokens`	integer	Sum of input (prompt) tokens.
`output_tokens`	integer	Sum of output (completion) tokens.
`total_tokens`	integer	`input_tokens + output_tokens`.
`cost_usd`	number	Total cost in USD for the group, rounded to 6 decimals.

[
  {
    "user_id": "3a1f8c22-9d4e-4b70-8c11-2f6a0e7b5d13",
    "user_email": "alex@example.com",
    "period": "2026-07",
    "provider": "anthropic",
    "model": "claude-opus-4-7",
    "byok": 0,
    "requests": 128,
    "input_tokens": 540200,
    "output_tokens": 96400,
    "total_tokens": 636600,
    "cost_usd": 3.11
  }
]

GET /cogs/per-agent

Cost and usage aggregated per agent × period (Zeitraum) × KI-Modell, ordered by cost descending. This is the per-agent cost-allocation view: it answers "which agent spent how much, on which model, in which period." Required role: admin or tenant_admin. It is the agent-dimension twin of /cogs/per-user — same period buckets, same tenant scoping, same row shape — with the grouping key swapped from the user to the agent.

curl "https://<your-gateway-host>/admin/v1/cogs/per-agent?range=30d&bucket=month"

Only agent-invoked traffic is included (requests that carry no agent are excluded). Agents are grouped by their clone-lineage root, so a shared agent and all of its clones roll up under one row group — matching the per-agent stats in /stats/analytics. Only genuine LLM legs are counted: legs with a NULL model (non-LLM side services such as PII scrubbing or guardrails) are excluded, so the per-model figures are clean. Customer-paid (BYOK) legs are kept on their own rows via the byok dimension and are never summed into platform cost. A single agent request that fans out into several same-model legs (e.g. a tool loop) counts as one request.

Cost and token figures are tenant-bounded (server-side tenant scope, fail-closed): a tenant_admin can only ever see their own tenant's agent spend; only an admin may scope to another tenant or read across all tenants.

Additional query parameters

Parameter	Type	Default	Description
`bucket`	string	`month`	Period granularity for the `period` column. One of `day`, `week`, `month`, `total`. Any other value is rejected with `400 { "error": "invalid bucket" }` (fail-closed allowlist — the value is never interpolated into SQL). `total` collapses the whole window into a single period labelled `"total"`.
`tenant_id`	string	—	Restrict to one tenant. Honoured only for callers with the `admin` role. Every other caller (including `tenant_admin`) is scoped to their own tenant server-side; a spoofed `tenant_id` is ignored, and a caller with no tenant is refused `403`. A caller can never read another tenant's cost rows.
`agent_id`	string	—	Restrict to a single agent, matched against the clone-lineage root id. Must be a single scalar value — a repeated param (`?agent_id=a&agent_id=b`) is rejected with `400 { "error": "agent_id must be a single value" }` (fail-closed; never bound into SQL). Combined with the mandatory tenant scope: an `agent_id` whose traffic belongs to another tenant returns an empty result.
`billable_only`	string	`1`	Include only billable legs. Set to `0` to also include non-billable legs (e.g. free-tier / self-hosted). Non-billable legs still obey the `model IS NOT NULL` rule.

Plus the shared range / from / to window parameters above.

Response

Array of rows. Returns 500 { "error": ... } on a storage failure, 400 { "error": "invalid bucket" } on a bad bucket.

Field	Type	Description
`agent_id`	string	Clone-lineage root id the legs were attributed to (the stamped root, falling back to the agent's own id for pre-lineage rows).
`agent`	string	Display name of the root agent (`agent.name`, falling back to its slug, then the raw root id if the agent was deleted). For a cross-tenant cloned agent the label may resolve to the source agent's name; the cost and token figures always count only the caller's own tenant's legs.
`period`	string	Period bucket. Format depends on `bucket`: `2026-07-15` (day), `2026-W29` (ISO year-week), `2026-07` (month), or `total`. Boundaries are computed in the gateway database session timezone (UTC on the standard deployment). Note: the `week` bucket uses the ISO week-numbering year, so late-December usage can fall in the following calendar year's `W01` (e.g. `2025-12-31` → `2026-W01`) — while `day`/`month` use the calendar year.
`provider`	string	Provider that served the model.
`model`	string	Model id (never null — non-model side-service legs are excluded).
`byok`	integer	`1` if the legs used a bring-your-own-key credential (customer-paid), `0` for platform-paid. Kept as a separate dimension so customer and platform cost never merge.
`requests`	integer	Distinct end-user requests in this group (`COUNT(DISTINCT request_log_id)`), not the leg count. This is a per-row count and is not additive across rows: a single request that fans out across models (a fallback) or across a period boundary is counted once in each row it touches, so summing the `requests` column overstates the true request total.
`input_tokens`	integer	Sum of input (prompt) tokens.
`output_tokens`	integer	Sum of output (completion) tokens.
`total_tokens`	integer	`input_tokens + output_tokens`.
`cost_usd`	number	Total cost in USD for the group, rounded to 6 decimals.

[
  {
    "agent_id": "ddc230d8-78ef-11f1-a3ae-6c92cf1260f0",
    "agent": "Support Triage",
    "period": "2026-07",
    "provider": "anthropic",
    "model": "claude-opus-4-7",
    "byok": 0,
    "requests": 64,
    "input_tokens": 210400,
    "output_tokens": 38800,
    "total_tokens": 249200,
    "cost_usd": 1.42
  }
]

Example curl requests

⭐ Example: The following examples show common query patterns for the timeseries endpoint.

Today's requests by hour (last 24 hours)

curl "https://<your-gateway-host>/admin/v1/stats/timeseries?bucket=1h&n=24"

Yesterday's requests by hour

# Set 'until' to the end of yesterday (start of today in Unix seconds)
YESTERDAY_END=$(date -d "today 00:00:00" +%s)
curl "https://<your-gateway-host>/admin/v1/stats/timeseries?bucket=1h&n=24&until=${YESTERDAY_END}"

Last 7 days by day

curl "https://<your-gateway-host>/admin/v1/stats/timeseries?bucket=1d&n=7"

Last hour in 5-minute buckets

curl "https://<your-gateway-host>/admin/v1/stats/timeseries?bucket=5m&n=12"

Last 30 days in 6-hour buckets

curl "https://<your-gateway-host>/admin/v1/stats/timeseries?bucket=6h&n=120"

GET /stats/reactivation

The platform-wide roll-up of the member-reactivation e-mails, shown as the Member reactivation card on the platform dashboard.

curl https://<your-gateway-host>/admin/v1/stats/reactivation

Authorization: platform admin only (cross-tenant data; 403 otherwise).

Response: mails_sent_7d, mails_sent_30d, reactivated_7d (members who became active within 7 days of a mail sent in the last 30 days), opt_outs_total, opt_outs_30d, and top_tenants — the ten most-mailed workspaces of the last 30 days as [{ tenant_id, slug, mails_sent_30d }] (an empty array when nothing was sent). Counts come from the send log only — no per-member classification runs fleet-wide; the per-tenant states live on GET /tenants/{id}/reactivation-metrics. A database fault is 503.

Stats API

GET /stats

Query parameters

Response structure

PeriodStats fields

TenantStats fields (GET /stats)

Example response

GET /stats/analytics

Query parameters

KI-Manager group scoping

Response structure

ChannelStats fields (UI vs API vs agent)

AgentStats fields (per agent, rolled up over the clone lineage)

AgentSeries fields (per-agent calls over time)

ActivityCounters fields

LatencyPercentiles fields

TopModelRow fields

GatewayStats fields

UserStats fields

CacheEfficiency fields

Example response

GET /stats/timeseries

Query parameters

TimeseriesPoint fields

Example response

GET /health/dashboard

Cost-of-goods (COGS)

Query parameters (all COGS endpoints)

CSV export (format) — cost-allocation endpoints

GET /cogs/per-model

Additional query parameters

Response

GET /cogs/per-tenant

Response

GET /cogs/per-service

Additional query parameters

Response

GET /cogs/per-token

Additional query parameters

Response

GET /cogs/per-user

Additional query parameters

Response

GET /cogs/per-agent

Additional query parameters

Response

Example curl requests

Today's requests by hour (last 24 hours)

Yesterday's requests by hour

Last 7 days by day

Last hour in 5-minute buckets

Last 30 days in 6-hour buckets

GET /stats/reactivation

See also

CSV export (`format`) — cost-allocation endpoints