Built-in dashboards

The extension plugs straight into Otoroshi's user dashboards. Turn on user analytics once, and your AI gateway comes with a full console: what it costs, who spends it, how fast and how reliably models answer, what your traffic emits, and what your agents call through MCP — no Kibana, no Grafana, no pipeline to maintain.

The overview answers the first questions in one screen: how much traffic, how much of it fails, what it costs, and which vendors and models carry it.

Every number is also a query you can drop into your own dashboards, filter by route, API, API key, group or tenant, and alert on: a daily spend threshold, an error rate on a provider, a tool that starts failing.

Getting started

Dashboards read from Otoroshi's user-analytics database (PostgreSQL). If you have not set it up yet:

1. Go to Data exporters and create a User Analytics (PostgreSQL) exporter.
2. Mark it as the active user-analytics exporter.

That's it. On the next sync the extension installs its dashboards next to Otoroshi's own, and every LLM call, MCP request and AI alert starts landing in its own tables.

MCP events are opt-in

MCP servers exposed by Otoroshi report their traffic when emit_audit_events is enabled on the MCP plugin or virtual server; MCP connectors report theirs when audit_events is enabled on the connector. Turn them on for the servers you want to see in the MCP dashboards.

The installed dashboards are ordinary user dashboards: rearrange them, change a widget, add your own queries, rename or delete them. A dashboard you have not touched stays up to date as the extension improves it; one you have customised is yours and is never overwritten. Delete it to get the latest version back on the next restart.

The dashboards

Dashboard	What it answers
AI Gateway - Overview	Calls, error rate, spend, tokens, vendors and models at a glance; usage by modality, weekly activity pattern, latest calls
AI Gateway - Costs	Spend over time and by model, provider, API key, user and route; projected monthly spend; the effective price per million tokens; a per-model summary table
AI Gateway - Tokens & cache	Input, output and reasoning tokens; who consumes them; prompt weight and output ratio per model; cache hit rate
AI Gateway - Budgets	Where every budget stands in its current cycle, spend and tokens per budget, budget warnings and overruns
AI Gateway - Ecological impact	Energy, CO2eq emissions (usage and embodied), water, primary energy and mineral depletion — and the carbon efficiency of each model
AI Gateway - Performance	Latency percentiles and heatmap, slowest models and providers, generation speed in tokens per second
AI Gateway - Reliability	Error rate, error kinds and messages, failing providers and models, guardrail denials, provider quota incidents, rate-limit headroom
AI Gateway - Consumers	The chargeback view: calls, tokens, spend, errors and emissions per API key, user and route
MCP - Overview	MCP traffic served and consumed, methods, transports, protocol versions, top and failing tools, servers and connectors
MCP - Tools & security	A per-tool table, tool latency, failing connectors and resource fetches, and every zero-trust event (rug-pulls, guardrails, redactions)

The costs dashboard: what was spent, what the month will cost at this pace, how the bill splits between input, output and reasoning tokens, and which models drive it.

Tokens & cache: where the tokens go — input versus output versus reasoning, model by model, provider by provider.

Numbers you can trust

• Every call is counted once. Behind a load balancer, a router or a fallback, the call is attributed to the provider that actually served it — spend and tokens are never doubled.
• Every modality is covered. Chat, responses, embeddings, images, audio, video, moderation and OCR share the same columns, so one widget compares them all.
• Cache hits are visible. They count as calls, with no tokens and no cost, and feed the cache hit rate.
• Units read without a calculator. Dollars, watt-hours, grams of CO2eq, litres of water, milligrams of antimony-equivalent.
• Content stays out. Prompts, completions, embeddings, images, transcripts, tool arguments and tool results are never written to the analytics database, and neither are API key secrets or user sessions. What is stored is what the dashboards read: who, which model, how much, how long, and whether it worked.

Queries

All queries appear in the widget wizard under their name. Scalar queries support the compare with previous period option.

Every LLM query also takes a modality parameter — chat, responses, completion, embedding, image, audio, video, moderation or ocr — to build the same view for one kind of usage: image generation spend, embedding latency, audio tokens by API key.

LLM usage

Query	Widget
cloudapim_llm_requests_total	metric	LLM calls
cloudapim_llm_errors_total / cloudapim_llm_error_rate	metric	Failed calls, and their share
cloudapim_llm_requests_over_time	area	Successful and failed calls
cloudapim_llm_requests_per_second	line	Call rate
cloudapim_llm_by_provider_kind / cloudapim_llm_by_provider	donut	Calls per vendor, per provider
cloudapim_llm_by_modality / cloudapim_llm_by_operation	donut / pie	Calls per modality, per endpoint family
cloudapim_llm_streaming_ratio	donut	Streaming versus blocking
cloudapim_llm_top_models / _top_providers / _top_apikeys / _top_users / _top_routes	bar	Rankings by calls
cloudapim_llm_requests_by_model_over_time / _requests_by_modality_over_time	line / area	One series per model, per modality
cloudapim_llm_activity_heatmap	heatmap	When the gateway is used: calls by weekday and hour
cloudapim_llm_distinct_users / _distinct_apikeys / _distinct_models / _distinct_providers	metric	How many consumers, models and providers are active
cloudapim_llm_apikeys_table / _users_table / _routes_table	table	Calls, tokens, spend, errors and emissions per consumer
cloudapim_llm_recent_calls	table	The latest calls: consumer, model, tokens, cost, duration, status

Tokens and cache

Query	Widget
cloudapim_llm_tokens_total / _input_tokens_total / _output_tokens_total / _reasoning_tokens_total	metric	Tokens consumed
cloudapim_llm_tokens_per_request	metric	Average tokens per call
cloudapim_llm_tokens_over_time	area	Input, output and reasoning tokens
cloudapim_llm_tokens_by_model / _by_provider / _by_modality / _by_apikey / _by_user / _by_route	bar / donut	Who and what consumes tokens
cloudapim_llm_prompt_completion_ratio	metric	Prompt tokens sent per token generated
cloudapim_llm_tokens_by_model_over_time	line	One series per model
cloudapim_llm_tokens_per_request_by_model	bar	Where prompts or answers are heaviest
cloudapim_llm_output_ratio_by_model	bar	Generated tokens per prompt token
cloudapim_llm_cache_hit_rate / _cache_hits_total	metric	Cache effectiveness
cloudapim_llm_cache_over_time / _cache_by_status	area / donut	Hits, misses, refreshes

Costs

Query	Widget
cloudapim_llm_cost_total / _cost_per_request	metric	Spend, and the average call
cloudapim_llm_input_cost_total / _output_cost_total / _reasoning_cost_total	metric	Spend by token type
cloudapim_llm_cost_per_1k_tokens	metric	The blended price paid per thousand tokens
cloudapim_llm_cumulative_cost_over_time	area	The burn curve: running total over the period
cloudapim_llm_cost_monthly_projection	metric	The selected period extrapolated to 30 days
cloudapim_llm_cost_over_time	area	Input, output and reasoning cost
cloudapim_llm_cost_by_model / _by_provider / _by_modality / _by_apikey / _by_user / _by_route	bar / donut	Where the money goes
cloudapim_llm_cost_by_model_over_time	line	One series per model
cloudapim_llm_cost_per_million_tokens_by_model	bar	The price actually paid, reasoning included
cloudapim_llm_cost_by_source	pie	Price table versus provider-reported costs
cloudapim_llm_unpriced_requests	metric	Calls with no cost attached — a model missing from the price table
cloudapim_llm_models_table	table	The model leaderboard: calls, tokens, spend, $ per 1k tokens, latency, error rate, gCO2eq per 1k tokens

Budgets

Query	Widget
cloudapim_llm_budgets_consumption	table	Spent and allowed dollars and tokens, per budget, for the current cycle
cloudapim_llm_budgets_usage	bar	How close each budget is to its limit
cloudapim_llm_requests_by_budget / _cost_by_budget / _tokens_by_budget	bar	Usage counted against each budget
cloudapim_llm_cost_by_budget_over_time	line	One series per budget
cloudapim_llm_budget_exceeded_total / _budget_almost_exceeded_total	metric	Overruns and warnings
cloudapim_llm_budget_alerts_over_time / _budget_alerts_by_budget	area / bar	Which budgets are under pressure, and when

The two budget status queries read the budgets' live counters — the ones that enforce the limits — so they are exact for the whole cycle, whatever period the dashboard shows.

Ecological impact

The ecological impact dashboard: energy, emissions and water behind the calls, split between usage and hardware manufacturing, and the models that weigh the most.

Query	Widget
cloudapim_llm_energy_total	metric	Energy, Wh
cloudapim_llm_gwp_total / _gwp_per_request / _gwp_per_1k_tokens	metric	Emissions, gCO2eq
cloudapim_llm_gwp_usage_total / _gwp_embodied_total	metric	Usage and embodied emissions
cloudapim_llm_cumulative_gwp_over_time	area	Running total of emissions
cloudapim_llm_pe_over_time	area	Primary energy
cloudapim_llm_wcf_total / _pe_total / _adpe_total	metric	Water (L), primary energy (MJ), abiotic depletion (mgSbeq)
cloudapim_llm_gwp_over_time	area	Usage and embodied emissions
cloudapim_llm_energy_over_time	area	Energy
cloudapim_llm_gwp_by_model / _by_provider / _by_apikey / _energy_by_model	bar / donut	Who and what emits
cloudapim_llm_gwp_by_model_over_time	line	One series per model
cloudapim_llm_gwp_per_1k_output_tokens_by_model	bar	Carbon efficiency: the same answer, a very different footprint

See Ecological impact for how impacts are estimated.

Performance and reliability

Performance: how long callers wait — median and tail — how fast tokens come out, and when latency spikes.

Reliability: the error rate over time, and straight to the cause — the actual provider messages, and the providers and models they come from.

Query	Widget
cloudapim_llm_latency_avg / _latency_p50 / _latency_p95 / _latency_p99	metric	Call duration
cloudapim_llm_latency_distribution	bar	Calls by duration band
cloudapim_llm_output_tokens_per_second / _output_tokens_per_second_over_time	metric / line	Generation speed
cloudapim_llm_latency_avg_over_time / _latency_percentiles_over_time	line	Average, p50, p95, p99
cloudapim_llm_latency_heatmap	heatmap	Durations by band over time
cloudapim_llm_latency_by_model / _latency_p95_by_model / _latency_by_provider	bar	Who makes callers wait
cloudapim_llm_output_tokens_per_second_by_model	bar	Generation speed
cloudapim_llm_error_rate_over_time	line	Error rate
cloudapim_llm_errors_by_kind / _top_error_messages	donut / bar	Status codes, exceptions, guardrail denials, open circuits
cloudapim_llm_errors_by_provider / _error_rate_by_provider / _errors_by_model / _errors_by_route	bar	Where failures come from
cloudapim_llm_recent_errors	table	The latest failures, with their error
cloudapim_llm_guardrail_denials_total / _guardrail_denials_over_time	metric / area	Calls refused by guardrails
cloudapim_llm_ratelimit_headroom_by_provider	bar	The providers closest to throttling you
cloudapim_llm_providers_table	table	Calls, error rate, average and p95 latency per provider
cloudapim_llm_provider_quota_alerts_total / _over_time / _by_provider	metric / area / donut	Provider quota incidents
cloudapim_llm_provider_quota_refused_calls	metric	Calls refused while a provider was throttled or out of credit

MCP

Query	Widget
cloudapim_mcp_calls_total / _tool_calls_total	metric	MCP requests, tool calls
cloudapim_mcp_errors_total / _error_rate	metric	Protocol, transport and tool failures
cloudapim_mcp_latency_p95	metric	Request duration
cloudapim_mcp_calls_over_time / _error_rate_over_time	area / line	Traffic and failures
cloudapim_mcp_by_method / _by_side / _by_transport / _by_protocol_version	donut / pie	How clients talk to your servers
cloudapim_mcp_top_tools / _failing_tools / _slowest_tools	bar	Tool rankings
cloudapim_mcp_calls_by_tool_over_time	line	One series per tool
cloudapim_mcp_tools_table	table	Calls, failures, average and p95 latency per tool
cloudapim_mcp_top_servers / _top_connectors / _failing_connectors	bar	Servers exposed, upstream servers consumed
cloudapim_mcp_top_apikeys / _top_users	bar	Who uses MCP
cloudapim_mcp_top_errors / _failing_resource_fetches	bar	What goes wrong
cloudapim_mcp_latency_percentiles_over_time / _latency_heatmap	line / heatmap	Latency
cloudapim_mcp_zero_trust_events_total / _zero_trust_blocked_total	metric	Zero-trust decisions
cloudapim_mcp_zero_trust_over_time / _by_kind / _by_tool	area / donut / bar	Rug-pulls, guardrails and redactions, tool by tool

Where the data lives

Three tables are created next to Otoroshi's own, and honour the exporter's retention settings:

Table	Holds
<table>_cloudapim_llm_usage	One row per LLM call: modality, provider, model, tokens, cost, ecological impacts, budgets, cache status, latency, error
<table>_cloudapim_mcp_calls	One row per MCP request, served or sent through a connector: method, tool, transport, protocol version, connector, latency, failure
<table>_cloudapim_ai_alerts	Budget warnings and overruns, provider quota incidents, MCP zero-trust events

Each row carries the Otoroshi request_id, so a call can be joined with the gateway event of the same request — and with the MCP tool calls it triggered.