Activity and conversations The audit trail, filters, risk findings, prompt capture. Overview Activity & conversations Activity is the audit trail: one row per AI request, recorded by the Gateway after the response streams back. What's recorded Model, tool (Claude Code / Cursor / …), developer identity, device, connection mode (subscription pass-through vs managed key), outcome (ok / blocked / upstream error), token counts, latency, MCP servers declared — and any risk findings embedded on the row (detector, action taken, severity). Request and response content is never stored by default. Prompt capture (opt-in) If your organization enables prompt capture, transcripts are stored redacted — secrets and personal data are replaced with [REDACTED:kind] markers before anything is written. Capture is off by default, and audit rows deep-link to the captured conversation when it exists. Blocked conversations When the AI classifier issues a blocking verdict, the whole conversation is flagged: subsequent requests are rejected until an admin clears the block from the Policy page. This turns a finding into an enforced stop, not just a log line. Compliance evidence The Compliance screen exports an evidence pack (policies in force, detector actions, activity summaries) for auditors — see the Compliance & SIEM guide. Read the activity feed Activity is the audit trail: every AI request your developers send through the Gateway, with the policy decision attached. The small dot beside the title means the view is updating live. The four tiles Requests and Blocked for the selected window, plus Top models and Top MCP servers — a quick read on what your team leans on. The columns Column What it tells you When Request time Developer Who, and which tool (Claude Code, Cursor, Copilot…) Model The model actually used Signals Risk findings and MCP servers involved Mode Subscription (their own plan) or API key (your managed key) Outcome OK, Blocked, or an upstream error Tokens / Duration Size and latency A blocked row is tinted and carries a red rail on the left, so it stands out when you're scanning rather than reading. Filters Time window, outcome, event type, developer, and a model filter. The common investigation — "show me everything blocked in the last 24 hours" — is two clicks. Mode is worth understanding Subscription means the developer's own AI subscription paid for that request and Sentilai only governed it. API key means it went through your organization's provider key and is billed to you. Both are audited identically; the difference is who pays. The Activity feed: every AI request with its policy decision. Blocked rows carry a red rail. Read a conversation When prompt capture is on, Conversations holds the transcripts. The list Every captured conversation with the developer, the tool, an excerpt of the first message, how many messages it has, and when it was last active. Filter by developer. The transcript Messages in order, each labelled by role — Developer, Assistant, or System prompt — with timestamps. A "truncated" badge means the content exceeded what is stored for a single message. Seeing the system prompt is often the most instructive part: it shows what the tool told the model before your developer typed anything. Redaction markers Amber shield badges naming a kind of secret mark where redaction happened. The transcript contains [REDACTED:aws_access_key_id] rather than the key, because redaction runs before storage — the value was never written down. Two different empty states "Prompt capture is off", with a link to Policy. Nothing is wrong; the feature is not on. "No conversations captured yet" — capture is on, nothing has come through. Telling these apart is deliberate. An admin who sees a blank list usually assumes a bug. What support can see Nothing here. Sentilai support can see that a request happened and what the policy decided; conversation content is not readable by us. Conversations — reconstructed sessions, grouped per developer. Secrets are redacted before storage, so a pasted key is already gone here. Filters and the four tiles Activity is a live feed — it refreshes every ten seconds, and the pulsing dot beside the title means what you are looking at is current. The tiles Over the selected window: Requests, Blocked, Top models (top three) and Top MCP servers (top three). Each shows "No data yet" rather than a zero when there is genuinely nothing, so you can tell "quiet" from "broken". Top MCP servers is the one people find most surprising. It is a ranking of what your AI assistants are actually reaching for. The filters Time window — last 24 hours, 7 days or 30 days. Outcome — all, OK, blocked, or upstream error. Event type — all, requests, or risk events. Developer — a specific person. The list includes admins as well as developers, because admins can have governed tools too. Model — free text, so opus narrows to everything with that in the model name. Combinations worth knowing Blocked, last 7 days — the review you should do weekly. It tells you whether your policy is doing something useful or just being annoying. Risk events, last 30 days — what the detectors are finding, independently of whether anything was stopped. One developer, 24 hours — what you look at when somebody says "it stopped working". The four tiles and the filter row. "No data yet" rather than a zero, so quiet reads differently from broken. The Signals column The densest column in the table, and the one that repays learning. MCP server badges Which MCP servers this request declared. This is how you find out that a request to summarize a pull request also had a database server and an HTTP client available. Risk finding badges One per detector that fired, in a warning colour, with a count. secrets ×2 means the secret scanner found two. The badge never shows the value. Findings carry the kind and the count only — the detectors are built so that the matched text is never stored. For the dependency detector, the tooltip does show the offending package names, because a package name is not a secret and knowing which one was proposed is the entire point. The classifier badge A millisecond figure — how long the prompt-injection classifier took — with the score in its tooltip. It is shown because the classifier runs in parallel with the upstream call and you should be able to confirm that it is not costing your developers latency. Reading a row quickly A row with no badges is ordinary traffic. A row with MCP badges and no risk badges is an agent doing agentic things safely. A row with several MCP badges and a risk finding is the one to open. The other columns When Request time. The feed is newest-first. Developer Who sent it, and which tool. A dash with a tooltip means the request arrived with a device token issued before attribution was in place — the developer should reconnect the tool in the Endpoint Suite, after which their name appears. Model The model actually used, which is not always the one the developer thinks they selected. Mode Subscription or API key — whose money paid. See Connection modes. Outcome OK, Blocked, or an upstream error passed through from the provider. Blocked rows are tinted and carry a red rail down the left edge, so a scan of the screen finds them without reading. Tokens and duration Size and latency. Useful for spotting the agent loop that is quietly burning ten times more tokens than anybody expected. Opening a row Clicking a row opens its captured conversation — but only when there is one. That requires prompt capture to have been on at the time. Without it, the row is the record.