About Keld Atlas
Keld Atlas is the web application where you analyze your organization's AI spend and configure your Keld workspace. Every Keld Atlas workspace belongs to one organization, and everything you see is scoped to it.
The sidebar has three parts. At the top, the Understand group has the analytics pages: Briefing, your dashboard of spend and usage, and Observability, where usage is attributed across teams, people, repositories, and models. People who are not administrators also see Connect tools here, for downloading Keld Signal. Below that, the Optimize group previews Job Categories, a ledger of spend by type of work. Administrators see a third group pinned to the bottom of the sidebar, with no heading of its own: Integrations, Classifications, Teams, and Activity. Settings sits below that again, in the account menu.
Roles at a glance
Keld Atlas has two roles, scoped to the workspace. Read the guide that matches your role: the Keld Atlas Administrator Guide or the Keld Atlas Viewer Guide.
| Role | Sees | Can do |
|---|---|---|
| Admin | Organization-wide Briefing, Spend, and Observability; the Activity feed; the Optimize preview; and the administrator-only pages | Connect provider accounts, manage teams, people, agents, and seats, create and edit Dimensions, export the Briefing, and set organization enrichment policy |
| Viewer | A personal Briefing covering their own AI usage, and the Optimize preview | Review their own usage and download Keld Signal from the Connect tools page. Cannot see organization-wide analytics or the Activity feed, connect provider accounts, manage teams, or create Dimensions |
Administrators connect AI provider accounts and manage tool telemetry from Integrations. The standalone Connect tools page is shown only to non-administrators, so Viewers can download Keld Signal without access to any administrator page.
Briefing
Briefing is the landing page, and what it shows depends on your role.
Administrators see an executive dashboard covering a selectable timeframe: a calendar quarter (the current quarter is shown quarter-to-date), the last 7 days, or the last 30 days. Its heading reflects the selected timeframe (for example, "30 Day Brief" or "Q3 2026 Brief"). You can filter by team and by provider. It is laid out in two columns, finance and engineering, and includes:
- Total inference spend, with a change against the prior period and a forecast to the end of the quarter
- Spend vs. budget or run rate, and spend over time combining usage and flat seat costs
- Capitalizable spend, the CapEx and OpEx split, shown as a strip beneath the spend chart
- Subscription vs. consumption split, configurable to any custom Dimension you have defined
- Attribution coverage, showing how much spend is unattributed and which members are untracked
- Breakdowns: spend by cost center and job category, top spenders, cost drivers by tool and model, activity mix, and premium-model share
- Recommended actions for reducing spend and closing attribution gaps
Cards link through to the Spend page for the underlying detail. Administrators can also export the Briefing as CSV.
Viewers see a personal Briefing covering only their own activity over the same timeframes: their spend or usage over time, a breakdown by work category and subcategory, and the models and tools they use. No other person's data is reachable from this view.
Spend
The Spend page holds the authoritative billed figures pulled from your provider account connections, alongside usage and cost by model, seat costs by provider, and the CapEx split. You can view it over the last 7, 30, or 90 days. Reach it from the linked cards on the Briefing.
Spend is distinct from the telemetry that drives the Activity feed and Observability, and Keld never merges the two. Billed figures come from your provider's admin API; telemetry cost is derived from token counts and is used for attribution.
Observability
Observability is where usage and cost are attributed across your organization. It covers the last 24 hours, 7 days, 30 days, or 90 days, and is organized into four tabs:
- Financial: attribution by cost center and financial category, including the built-in CapEx/OpEx Dimension
- Technical: attribution by model, repository, and telemetry source
- Team: attribution by team and by person
- Your Reports: any custom label Dimensions your organization has defined
Administrators create and edit Dimensions and their label rules here, using a rule builder that previews each rule against live data before you publish it. The CapEx/OpEx Dimension is built in: you can edit its rules or reset it to the defaults, but you cannot delete it.
Enrichments
Enrichments are the labels Keld Signal attaches to each prompt on the user's own computer: small, structured tags that turn raw AI usage into something you can attribute, group, and report on. Prompts and code never leave the machine, and only the labels reach Keld Atlas.
Each enrichment is one classification pass over the prompt. A pass picks a single label, applies any labels that fit, extracts spans of a given type (masked on the device), or pulls several related fields into one structured record. Passes run as a short pipeline, so a later pass can build on an earlier one: which Job Subcategory options apply depends on the Job Category decided first.
Keld ships built-in passes for task type, sensitivity, domain, activity, speech act, Job Category, and Job Subcategory; these are fixed and cannot be edited. Administrators add custom passes, and browse a gallery of ready-made ones, from Classifications. Custom definitions publish to every device in the organization. See Classifications in the Keld Atlas Administrator Guide.
In Keld Atlas, these labels drive the Job Categories view, the built-in CapEx / OpEx Dimension, and the sensitivity flags on Activity entries. Attribution to a person, team, or repository comes from the identity on each telemetry event rather than from enrichment. For the full mechanics, see How enrichments work in the Keld Signal Overview.
Activity
Activity is a live stream of your organization's AI tool telemetry, available to administrators. Events are grouped into turns: a prompt, the model's response, and any background calls the tool made on its own. Each turn is collapsible and carries a cost and token summary, with a raw-event inspector for the underlying data. All prompt and response content is redacted.
A configurable KPI strip sits above the feed, and you can filter by tool, model, person, team, repository, and agent. A work filter separates autonomous agent runs from human-initiated work, so you can look at either in isolation.
Agents & Apps
Autonomous agents are tracked as their own class of user. Go to Teams and open the Agents & Apps tab to see the agent roster, with headline figures for agent types, agents currently live, month-to-date spend, and total runs.
Agents appear in one of two states. Claimed agents are registered with their own identity and API key, which you manage from the Agent API Keys drawer on the Teams page. Discovered agents are ones Keld has seen in telemetry riding on a person's token but that have not yet been registered. Claiming an agent gives its spend a stable identity of its own rather than attributing it to the person who ran it.
Selecting an agent type opens a detail view of its runs. Agent spend also appears in the Activity feed, where the work filter separates it from human activity.
Keld Relay (preview)
Keld Relay is a forward-looking job routing service, shown in the sidebar under Optimize. Job Categories is the one page enabled today: a ledger that classifies work by business function and estimates what routing each category to the Keld Exchange could save. Models, Model collections, and Deadline routing are coming soon.
Coming soon
Budgets and Billing are not yet available. Budget figures shown on the Briefing compare against a run-rate baseline until Budgets ships.
Data handling and privacy
This section applies to everyone who uses Keld. Keld Atlas collects usage metadata: model, token counts, latency, cost, tool, repository, team, person, and similar fields. It gathers this through provider account connections, tool telemetry, and Keld Signal. It does not collect or store prompt content.
- Telemetry configurations generated by Keld keep prompt logging off.
- Keld Signal classifies prompts on the user's computer and publishes only masked labels and masked entity spans. Raw prompt text never leaves the user's computer.
- Every query is scoped to your organization. Data from one organization is never reachable from another.
Raw telemetry is retained for up to 90 days. Older partitions are archived to cold cloud storage before they age out, and rollups power the longer-term reporting views.