Dashboard

The Dashboard is the administrator's home page and the first entry of the sidebar. It answers three questions at a glance: how many users, devices and AI agents the account has, how the company's AI agents use AI on devices and through the Netzilo AI Gateway, and which operating systems and security scores the devices have.

Only administrators and owners see the Dashboard. Regular users land on the Workplace.

The Dashboard with the tiles Users 13, Devices 8, Agents 21 and Tokens 9,000, each with a seven-day line, and below them Agentic Activity set to Last 24 Hours with its Refresh and View buttons: the cards Tokens, Cost and MCP Calls, a Top Teams bar graph with five icons to pick the list, and a Tokens chart with three icons to pick the figure

How it works

The Dashboard shows a snapshot that Netzilo Server computes once an hour; the Agentic Activity window shows views it builds every 15 minutes. A request made a minute ago is on neither until they are built again. To build both now, click Refresh (the circular arrows) on the Agentic Activity panel or in the window. For live numbers, generate an AI Usage report.

The window's Sessions tab is the exception: it reads the sessions live and reads them again every 30 seconds while it is open.

Tiles

Four tiles sit at the top. Each shows a total and a line for the last seven days; hover over a point to see that day's value.

TileCounts
UsersUsers of the account
DevicesDevices (peers) of the account
AgentsDistinct AI agents active in the last 7 days
TokensTokens recorded per day, by the AI Gateway and by Netzilo clients that record AI sessions

Agentic Activity

Agentic Activity summarises the AI sessions recorded on devices and on the AI Gateway. Choose the period with Last 24 Hours, Last 7 Days or Last 30 Days at the top right. Next to it, Refresh builds the numbers now and View (the two arrows) opens the Agentic Activity window.

The panel shows three cards, a top list and a chart:

  • Tokens (tokens in and out), Cost (and the cost per request) and MCP Calls (with errors and blocked calls).
  • The top list, as a bar graph. Its icons pick the list: Top Teams, Top Guardrails, Top Models, Top Agents and Top Tools. Hover over an icon for its name.
  • The chart over the period. Its icons pick the figure: Tokens, LLM Requests and MCP Calls.

Agentic Activity for Last 7 Days with Top Teams listing Engineering, Finance, Developers and Sales by tokens, and the chart switched to LLM Requests with a bar per day

Where the numbers come from:

  • The cards and the chart count the recorded usage: the model calls and MCP calls that the AI Gateway and the Netzilo clients recorded. Browser traffic and older clients are not in them. Cost is computed for gateway traffic; calls recorded on a device count tokens at a cost of $0.
  • Top Agents, Top Models and Top Tools count the calls recorded in AI sessions from every source: the Netzilo client on devices, coding-agent hooks, the browser extension and the AI Gateway. They show each one's share of the total.
  • Top Guardrails counts the guardrail verdicts (AI Edge events) by rule, as shares.
  • Top Teams ranks the users' groups by tokens.
  • A top list shows its five largest rows.

Before there is any agent activity, the section shows Getting Started with AI Edge instead, with a Get Started with AI Edge button that opens Edge > Providers. See AI Gateway and Providers.

Refresh

Refresh builds the snapshot and the window's views now, so a request made a moment ago shows up. It takes a few seconds, longer on a busy account, and the arrows turn until it is done. Netzilo Server builds an account's numbers once at a time and at most once every 30 seconds: when several administrators click Refresh together, they all get the same build, and a click within 30 seconds of the last build shows that build.

Agentic Activity window

The View button (the two arrows) on the Agentic Activity panel opens the Agentic Activity window. It goes from the whole company down to a single agent session: the teams, a team's people, agents, models, MCP clients, servers and tools, the sessions behind each of them, and what happened in a session step by step. The sessions are those recorded on devices and on the AI Gateway.

To find a session without browsing teams, use AI Sessions under Activity: it lists every recorded session and opens it in this window.

The first page has three tabs, Overview, Teams and Sessions, with the period on the right of the tabs. Overview is the panel of the home page, filling the window.

The Agentic Activity window on the Overview tab for Last 7 Days: the cards Tokens 23M, Cost $5.09 and MCP Calls 4, Top Teams with Engineering, Finance, Developers and Sales, and a Tokens chart with a bar per day

The window can be dragged, resized from any edge and maximized with the button next to Refresh. Its title names where you are, with that level's icon. The arrow at the top left, or Escape, goes back one level; Escape on the first page closes the window.

Teams tab

Teams shows a tile for each group whose members had recorded AI sessions in the period. All counts everyone.

  • A tile shows the team's name, how many agents, people and service users were active, and its tokens over the period as bars. Hover over a bar for the tokens and cost of that hour or day.
  • The colored bar on the left of a tile is the team's tokens against the busiest team's, from gray (few) to blue (the busiest team), as the Tokens legend shows.
  • Tiles are ordered by tokens, the busiest first. More tiles load as you scroll down. Search teams finds teams by name among all of them, not only those loaded.
  • Hover over a tile for its sessions, tokens, cost and violations. Violations count what the guardrails blocked, redacted or detected.

The Teams tab for Last 7 Days: a Tokens Low to High legend, a Search teams field, and tiles for All, Evaluation, Engineering, Finance, Developers, Sales and other teams, each with its agents, people and token bars

Click a tile to open the team.

A team's people, agents and models

A team opens as a flow. Traffic at the top switches between two flows:

  • LLM: from the Users through their Agents to the Models, grouped by provider. The width of a band is the tokens that went that way. Hover over a name or a band for its tokens, cost and violations.
  • MCP: from the Users through the MCP Clients and MCP Servers to the Tools. The width of a band is the number of calls. The MCP flow shows no verdicts.

The Developers team on the LLM flow for Last 7 Days: users on the left, agents such as Claude Code, OpenAI SDK (Python) and AWS SDK in the middle, models by provider on the right, with the Users, Agents and Models menus and the LLM and MCP switch at the top

The same team on the MCP flow: users on the left, then MCP clients, MCP servers and tools, with bands sized by the number of calls

A team with many people shows its 40 busiest at first. The others are one node at the bottom of the Users column, N more people. Scroll to the end of the flow, or click that node, to add the next 40, up to 500 people.

Users, Agents and Models narrow the LLM flow; Users, Clients, Servers and Tools (grouped by server) narrow the MCP flow. The period menu changes the period. Click a name to list its sessions.

Sessions

A session is one run of an agent: a Claude Code conversation, an SDK agent's task, the calls a client made without a session of its own. The list shows, most recently active first, when each session started and was last active, its agent and user, its status, the models it used, its turns with its model and tool calls, its tokens and cost, and its violations. A session with a kill chain carries a kill-chain badge that opens it on the behavior graph. Search finds sessions by name, agent or model.

Opened from the MCP flow, the list shows when each session started and was last active, its user and agent, its MCP servers, its calls and distinct tools, the average latency of a call, its violations, and Model + MCP when the session also called a model or MCP only when it did not.

The list is read a page at a time. The arrows at the bottom move between pages and show the page you are on, for example 1 of 456. Started, Last Activity, Turns, Usage and Violations (and Calls on the MCP list) sort every session of the list, not only the page shown.

The sessions of Alice Morgan for Last 7 Days: rows of Build Agent sessions with their status, start and last activity, model, turns with model and tool calls, tokens and cost, and at the bottom the page arrows showing 1 of 456

Click a session to open it.

Sessions tab

Sessions shows the sessions of the period as a honeycomb: one hexagon per session, grouped in a comb per team. A hexagon's color is its session's state, and nothing else:

ColorState
RedUnder Isolation: a response action isolated the session
OrangeUnder Inspection: a response action placed the session under inspection
GreenActive: the session is running
BlueIdle: the session is idle or finished. The more tokens it used, the brighter the blue: under 1K, 1K to 10K, 10K to 100K, 100K to 1M, and over 1M tokens

The Sessions tab for Last 7 Days: the legend Under Isolation 1, Under Inspection 1, Active 1 and Idle 13.5k, a Search teams field, and a comb per team, Finance with a red hexagon, Developers with an orange and a green one, Evaluation, No group, Engineering and Incident Response in shades of blue, with the Developers comb hovered: 1 person, Under Inspection 1, Active 1, Idle 11.4k, Tokens 1.3M, Shown 250 / 11.4k

  • A comb fills from its middle out: the isolated sessions first, then those under inspection, then the active ones, then the idle ones, those with the most tokens first.
  • A comb shows at most 250 hexagons, fewer when there are many combs. The number next to its name is always the exact count of its sessions; hover over a comb for its counts by state, its tokens, and how many of its sessions are shown.
  • The legend at the top counts the sessions of the page's level by state: every team's, or every one of the team's or the person's, whether its comb is shown or not. Click a state to show only its sessions; click more states to add them, and click again to remove one.
  • A person is in one team: the most specific group (the one with the fewest people) of the groups their latest session in the period recorded. People in no group but All are in No group.
  • With many combs, a name shows over each comb that is wide enough for it, or that holds isolated or inspected sessions; hover over a comb to see the others.
  • With more than 100 combs, or more than fit, the combs keep a small size and the page scrolls; more combs load as you scroll down. Search teams finds teams by name among all of them.

Click a team's comb, or its name, to open its people: a comb per person, with the legend and Search people at the top of the window.

The Evaluation team's people for Last 7 Days: the legend in the window's toolbar with Idle 1,860, a Search people field, and a comb per person, each named with its number of sessions

Click a person's comb to open their agents: a comb per agent, each hexagon one of the agent's sessions. Hover over a hexagon for the session's state, the reason a response action gave, its model, tokens and start time. Search agents narrows the combs.

Alice Morgan's agents for Last 7 Days: combs for Release Bot, Go-http-client, Python-urllib, Claude Code, Build Agent and others, with an orange hexagon of Release Bot hovered: State Under Inspection, Reason Instructions injected through a fetched page, Model claude-sonnet-4-6, Tokens 1,500, Started 3 minutes ago

Click a hexagon to open its session. The arrow at the top left, or Escape, goes back one level: from the session to the person, to the team, to the tabs.

A session

A session opens on its Trajectory. The tabs at the top right are Trajectory, Behavior Graph, which shows the same session as a graph, and, when the session made MCP calls, MCP Calls.

The Trajectory shows the session step by step:

  • A timeline with the lanes Input, Model, Tools, Host when the device reported activity, and Kill chain when the session has one. Duration, Turns and Calls change its scale.
  • Below it, every message in order, by turn: the system prompt, the user's prompts, the agent's context, the model's thinking and answers, and each tool call with its result. Guardrail verdicts and kill-chain steps are marked on the messages they fell on.
  • Long texts show their beginning; click a message for the whole text and its details. For a tool call, the Source tab names the server, the tool, who ran it, the MCP client, the MCP session and Seen at, the gateway or device that recorded it. For a model call, the Source tab shows the Request, the Model that answered, Asked for, the model the caller named, Routed by, the routing rule that sent the call elsewhere (A deleted routing rule once it is deleted), the Provider, the Step, Tokens, Cost, Duration and Started.
  • On a device, the agent's Host activity is shown under the tool call that caused it: the processes it started, the files it read, wrote, created, deleted or renamed, and the connections and requests it made. Activity that no tool call explains follows the step it came after. The summary counts processes started, files read and written, and connections.

The Trajectory of an agent session recorded on a device: 1 turn, 2 model calls, 2 tool calls, 3 processes started and 3 files read, a timeline with the lanes Input, Model, Tools and Host, the messages by turn, and the run_shell tool call selected with its arguments, result and Host activity: the shell, cat and curl started, the files they read and the request to example.com

The messages are those the AI Gateway or the device's Netzilo client recorded: what the agent sent and received, after the guardrails. Text a guardrail redacted is shown redacted, and a tool call the guardrails held back is not kept. See AI Gateway.

The Behavior Graph shows who and what the session touched: the user, the agent, the models, the tools and the device, with the steps between them. File activity is one node per program. Kill chains picks a detected chain and marks its steps; Filters hides node and edge types.

MCP Calls lists the session's MCP tool calls: each call's tool, its arguments and the start of its result, its verdicts (Redacted, Blocked, a kill-chain stage such as Execution) and its time. Click a call for its details, arguments and result.

A session on the MCP Calls tab: three calls of a lookup tool, redacted, redacted and blocked, and the first call selected with its server, status, verdicts, arguments and redacted result

The Behavior Graph of the same session: Alice Morgan runs Claude Code, which sends LLM requests to claude-haiku-4-5 and calls the Read tool, with the detected kill chain marked in red

Devices

The Devices panel at the bottom shows two charts:

  • OS Distribution: how many devices run Windows, OSX, Linux, iOS and Android.
  • Security Score Distribution: how many devices fall into each security score, from A (best) to F.

The Devices panel with an OS Distribution bar chart for Windows, OSX, Linux, iOS and Android, and a Security Score Distribution radar chart with the axes A to F

The security score of each device is on its row under Endpoint > Peers. See Peers.

When the account has no device yet, the panel is replaced by Get Started with Netzilo with an Add Peer button. See Install Netzilo.

Read the Dashboard through the API

These calls need an administrator's token. See Agents for tokens.

CallReturns
GET /api/statsThe saved snapshot, or null before the first hourly run
GET /api/ai/activity?window=24h&scope=orgThe window's first page: the totals and one tile per team. window is 24h, 7d or 30d
GET /api/ai/activity?window=24h&scope=org&offset=0&limit=60&q=salesA page of the tiles, the most tokens first, of those whose name holds q: limit tiles (at most 200) from offset. total_tiles is how many match and max_tokens the busiest team's tokens
GET /api/ai/activity?window=24h&scope=group:<group-id>A team's LLM flow from its users to agents and models, and its MCP flow, mcp_flow, from its users to MCP clients, servers and tools
GET /api/ai/activity?window=24h&scope=group:<group-id>&people=40The same flows over the team's people busiest people (at most 500). The others are one node with the key *others, its others field their number; each flow's people is how many people it has
GET /api/ai/hive?window=24h&offset=0&combs=100A page of the Sessions tab: counts, the sessions by state (isolated, inspected, active, idle, and idle_bands by tokens), a page of combs (at most 200) from offset, and total_combs. Add team=<group-id> for a team's people, user=<user-id> for a person's agents with their sessions (limit sessions per agent, at most 500), and q to search by name
GET /api/ai/sessionsThe sessions, filtered, searched and paged; see AI Sessions
POST /api/ai/activity/refreshBuilds the snapshot and the views now, as Refresh does

The GET answers carry an ETag: send it back in If-None-Match to get a 304 Not Modified when nothing changed.

Troubleshooting

Agentic Activity is empty although the AI Gateway works or devices record sessions. The snapshot is computed once an hour. Click Refresh, or generate an AI Usage report, which reads live data.

A team is missing from the window. A team appears when its members had recorded AI sessions in the period. Choose a longer period. A user who is in no group counts only under All.

A session has no messages. Only sessions recorded by the AI Gateway or by a Netzilo client that records AI sessions have messages; browser traffic has none. A call whose content was not captured shows Content not captured for this call.

The Tokens tile shows zero. Nothing was recorded that day. Tokens count the gateway's traffic and the model calls Netzilo clients recorded; browser traffic does not count.

The Sessions tab and the Teams tab count differently. Teams shows the views built every 15 minutes; Sessions reads the sessions live. A team's comb holds the people whose latest session recorded that team as their most specific group, so a person in several groups is in one comb only, while each of their teams' tiles counts them.

A comb shows fewer hexagons than its number. A comb shows at most 250 hexagons, fewer when the page has many combs, in the order isolated, inspected, active, idle. The number next to its name, and its counts on hover, are exact.

The Sessions tab shows Could not load. Netzilo Server builds a few session hives at once for all accounts, so that they never hold up the rest of the server; one that waits too long answers that it is busy. Click Retry, or wait: an open tab reads the hive again on its own every 30 seconds and keeps what it shows until then.

A flow ends with N more people. A team's flow shows its 40 busiest people at first. Scroll to the end of the flow, or click the node, to add more, up to 500.

Where is the Security Score? The Dashboard no longer has a score tile. The Devices panel shows how the scores are distributed, and Endpoint > Peers shows each device's score. Fix a low score by acting on the device's red indicators, or enforce requirements with Posture Checks.