Security dashboard
The Security view in your workspace gives you real-time visibility into every outbound HTTP request your agents make — what was allowed, what was blocked, how the AI judge decided, and which agents have no access policy at all.
Overview
Envelope's security proxy sits between your deployed agents and the internet. Every outbound request from an agent passes through the proxy, which evaluates it against the agent's access policy before forwarding. The result — and the reasoning — is recorded and visible in your workspace.
Open your workspace and click Security in the left sidebar (alongside Library and Observability). You'll see a live overview of every outbound call across all your installed teams.
Room thread — You can also surface security data directly in conversation. In any team's room thread, ask the assistant about recent outbound traffic, blocked requests, or policy coverage — it pulls from the same data without you needing to open the Security view.
Stat cards
Four summary cards appear at the top of the Security view, covering all your installed teams combined:
- Total Requests — every proxied outbound call across all teams
- Blocked — requests that matched a deny rule and were stopped (shows "no violations" when zero)
- Passthrough — requests from agents with no access policy, forwarded without any check (shows "all inspected" when zero)
- Policy Coverage — the percentage of agents across all teams that have an access policy declared, shown as N of M agents
The top three cards (Total, Blocked, Passthrough) refresh automatically every 30 seconds. Policy coverage refreshes when you reload the view.
By team
Below the stat cards, each of your installed teams appears as a row. Each row shows:
- A health dot — green if traffic is clean, amber if there are passthrough requests, red if requests have been blocked
- The team name
- Count badges for total requests, blocked, passthrough, and policy coverage for that team
Expanding a team row opens three sections inline:
Traffic audit
A table of every outbound request that team's agents have made, newest first. Each row shows the timestamp, decision badge, hostname, and agent name. Click any row to expand it and see the full URL, HTTP method, run ID, decision source, and — for JUDGE decisions — the AI's reasoning and confidence level.
Use the filter chips to narrow the view: All, Blocked, Judged, or Passthrough. Only filters with at least one matching row are shown. Up to 30 rows are displayed; a count appears if there are more.
External hosts
A breakdown of every external hostname the team's agents have called — how many times, which agents, and the outcome mix (allowed, blocked, judged, passthrough) for each host.
Access policies
A per-agent list showing whether each agent has an access policy declared. Agents with no policy show a No policy badge — these are the ones causing passthrough traffic. Agents with a policy show whether they default to Deny-default (safer) or Allow-default, along with up to four of their individual rules.
Alert events
If any teams have produced blocked requests, passthrough agents, or judge decisions, an Alert events feed appears below the team list. It collects these events across all your teams in one place, newest first.
All judge decisions appear in the feed. Low-confidence judge outcomes are flagged with a low confidence badge. Events that triggered an email alert are marked alert sent.
Events that send email alerts:
- Blocked requests — at most one email per team per hour
- New passthrough agent — once per agent (first time an agent runs without any access policy)
Judge decisions appear in the feed for visibility but do not send an alert email.
Decision types
Every request receives one of four decisions, shown as a coloured badge throughout the view:
ALLOW (< 2ms) — The request matched an allow rule in the agent's access policy and was forwarded. This is the expected path for legitimate agent calls.
BLOCKED (< 2ms) — The request matched a deny rule or called a host not covered by the access policy. The agent received a 403 response and no outbound request was made.
JUDGE (~1–3s) — The request matched a rule configured to require AI review. The proxy forwarded it to an AI model, which read the agent's role, the request details, and the rule's stated reason, then decided whether to allow or block. The reasoning is visible when you expand the row. A "low confidence" badge means the judge made a call it wasn't sure about.
PASS — The agent has no access policy declared at all. The proxy forwarded the request without any policy check. Agents showing this need a policy added to their team definition to get full enforcement.