Request logs

See every request and tool call through the AI Gateway, and decide whether Altrum keeps their content.

What Altrum records

For every request and every MCP tool call:

• when it happened, the key, the AI system and the person behind a personal key,
• the client that sent it, read from what the tool calls itself (its User-Agent): for example Claude Code, Cursor, Codex CLI or API. Other is a client Altrum does not recognise, and Unknown a request that named none,
• the provider, the model, and any fallback used,
• the tokens and the cost,
• how long it took,
• the outcome, one of the words in the table below,
• which guardrails checked it and what they found: the guardrail, the direction, the action, and the kind of thing found. Never the matched text.

AI Gateway, Request log shows your whole organization’s traffic, newest first, on two tabs: Requests and Tool calls. Each AI system’s AI Gateway tab shows its own requests under Requests. Filter Requests by date range, project, AI system, key, model, outcome, client, person and guardrail: Guardrail shows only the requests that guardrail fired on. Filter Tool calls by date range, project, AI system, tool, MCP server, person and guardrail. Project appears once your organization has a project. A row in a Reporting ranking opens this page with its filter already set. The Client column shows the client; hover it to see the full User-Agent.

Select Details on a request to see its provider, model, client and User-Agent, tokens, cost, latency, any fallback, and whether the prompt and the reply passed the guardrails. Under Guardrails it lists each guardrail that fired, with what it checked (Prompt or Reply), what it did (Blocked, Masked or Logged only) and, when the guardrail says, what it found, such as Email address. Where content was kept, owners, admins and Governance Leads select View content to read the prompt and the reply on Prompt and Reply tabs, in the same panel. Add to incident, at the foot of the panel, links the request to an open incident or opens a new one with it.

Health

AI Gateway, Overview shows, under the connection details, whether the AI Gateway is answering well over the date range. It opens on Today. What the traffic cost is in Reporting.

• Figures: Requests, Success rate, Median latency, 95th percentile latency, Error rate and Cache hit rate, each with its change against the same length of time just before.
• Requests charts requests by result (succeeded, blocked, failed), and Latency the median and 95th percentile.
• Providers lists each provider with its requests, error rate, failed requests and average latency. The error rate is marked green up to 2%, amber above 2% and red above 10%.

A request counts as succeeded when it was answered, from the model or the cache, including with text masked. It counts as blocked when Altrum refused it on purpose: a guardrail, a budget, a rate limit, or what the key may use. Everything else failed. Times are grouped by hour or day in UTC.

Date range, filters and sharing

• Pick a date range from Today, 7 days, 30 days, This month, Last month and This quarter, or pick a range of your own on the calendar: the first day, then the last, then Apply. Ranges are whole days that start at midnight in your browser’s time zone, up to 92 days. Pages open on 30 days.
• The range and the filters are part of the page’s address, so a link you share opens the same view.
• On a phone the filters are behind Filters, which shows how many are set.
• Export CSV saves the requests loaded in the log. A cell that starts with =, +, - or @ is saved with an apostrophe in front, so a spreadsheet does not run it as a formula. Content is never exported.
• Print prints the page in black on white, without the navigation.
• Each health figure shows its change against the same length of time just before: green when it is better, red when it is worse, grey for figures with no better direction, such as requests.

Outcomes

Outcome	Shown as	What happened
ok	Succeeded	The request was answered.
cache_hit	From the cache	The answer came from the cache.
masked	Masked	A guardrail masked part of the request or the answer, and the rest went through.
blocked_input	Blocked: prompt	A guardrail blocked the request. It was not sent to the model.
blocked_output	Blocked: reply	A guardrail blocked the answer. It was not returned.
guardrail_check_failed	Guardrail check failed	A guardrail check could not be completed, so the request or the answer was held back.
budget_exceeded	Over budget	A budget was used up.
rate_limited	Rate limited	A rate limit was reached.
model_blocked	Model not allowed	The key may not use that model, provider or tool, or your organization serves approved models only and the model is not approved.
key_refused	Key refused	The key may not make this request, or its organization or AI system is switched off.
provider_error	Provider error	The model provider returned an error.
provider_timeout	Provider timed out	The model provider took too long.
client_aborted	Caller stopped	The caller stopped waiting.
invalid_request	Invalid request	The request could not be understood or named a model that does not exist.
gateway_error	AI Gateway error	The AI Gateway failed to handle the request.

Content is off by default

Altrum does not keep what was sent or answered unless your organization chooses to.

• An owner or admin switches it on in AI Gateway, Settings, Content logging: Keep prompts and replies, for the organization’s keys. Each AI system can choose differently on its AI Gateway tab under Content logging: Follow the organization, On or Off.
• Kept content is the content after masking: text a guardrail masked is never stored.
• The request record itself never holds the content. It stays apart, under your organization only, and a record says whether content was kept for it.
• Content is kept for 30 days by default. Set Keep for (days) to any number from 1 to 365. When the time is up the content is deleted and the record says none is kept.
• Only owners, admins and Governance Leads can read stored content, inside Altrum, and every read is written to your organization’s audit trail.

How long records are kept

Request records without content are kept for 13 months, so you can compare a month with the same month a year earlier. Tool call records and guardrail findings are kept for the same time.