Introduction to Keld
Keld enables you to monitor and understand how your organization uses AI. You connect your AI provider accounts to Keld Atlas and install Keld Signal on your team's computers, then analyze spend by team, person, repository, model, and cost center in the Keld Atlas dashboard. Keld will later expand into a marketplace called Keld Exchange, where enterprises buy inference from vetted providers.
What you can do with Keld today
Keld collects spend data from two places, and you can use either one on its own:
- Provider account connections. Connect your Anthropic and OpenAI accounts to Keld Atlas to pull authoritative billed spend and usage over each provider's read-only admin API. Google Gemini spend is estimated from Gemini CLI telemetry; AWS Bedrock support is coming soon.
- Keld Signal on users' computers. Keld Signal adds tool telemetry and on-device enrichment, which is what makes spend attributable to a specific person, team, and repository.
With both connected, you can:
- Break down spend by tool, team, person, repository, cost center, model, business function, and financial category such as CapEx vs. OpEx.
- Watch AI activity from across your organization stream in live in the Activity feed.
- Create custom Dimensions to slice spend in a way that suits your organization and its processes.
- Export usage and spend data as CSV files for import into other financial analysis tools.
Keld Atlas
Keld Atlas is the main web application and user dashboard. It is the single interface where organizations analyze and manage their AI spend and identify non-urgent work that can save money if routed through Keld Exchange when the Exchange becomes available. Keld Atlas hosts usage observability, the Keld Relay interface (coming soon), and the administrator-only pages for workspace configuration.
Keld Signal
Keld Signal is the AI usage data layer of Keld. It securely pulls telemetry data from your team's AI tools (Claude Code, Codex, and Gemini CLI today, with Cursor, Antigravity, and agent frameworks such as LangChain and Vertex AI coming soon) and categorizes that usage on the user's computer. The aggregated data enables you to break down and view spend in the Keld Atlas dashboard by team, person, repository, model, cost center, and any custom Dimension you define, including CapEx vs. OpEx classification.
Important: Keld Signal performs all data enrichment on the user's computer. Keld does not transmit, store, or display your users' prompts anywhere in its systems.
Keld Signal installers are available for Windows, macOS, and Linux. You can download them from the Keld Atlas dashboard or during onboarding.
About the preview
Keld Atlas and Keld Signal are available as previews. Expect the following during the preview period:
- Available now: Keld Atlas (the Briefing, Activity, and Observability views, plus the administrator configuration pages) and Keld Signal for Windows, macOS, and Linux.
- Not yet available: Budgets, Billing, Keld Relay, Keld Exchange, and Keld Trade.
- Integrations coming soon: AWS Bedrock provider connections; Keld Signal support for Cursor and Antigravity; and Keld Signal support for agent frameworks including LangChain and Vertex AI.
- Changes to expect: interfaces and data models may change between preview releases, and some changes may require you to update Keld Signal.
To report a problem or send feedback, click the feedback icon in the bottom right of the Keld Atlas dashboard.
Before you begin
Setting up Keld involves two kinds of work. Check that you have what each one needs.
To set up an organization, you need:
- An admin API key from Anthropic (starts with
sk-ant-admin…, with usage read scope) or OpenAI (starts withsk-admin-…, with usage and billing read scope). These keys are usually held by whoever administers your provider account, so identify that person before you start. - The email addresses of the people you want to invite.
To connect your own computer, you need:
- An invitation to your organization's Keld Atlas workspace.
- Windows, macOS, or Linux.
- At least one supported AI tool installed: Claude Code, Codex, or Gemini CLI. Support for Cursor, Antigravity, and agent frameworks including LangChain and Vertex AI is coming soon.
Quickstart
There are two setup paths, and most organizations need both. An administrator sets up the organization once. Every person whose AI usage you want to measure then connects their own computer.
Set up your organization
Complete these steps once per organization. To set up a new Keld organization, follow these steps:
- Create your Keld account
Go to atlas.keld.co and sign up. A short onboarding wizard walks you through setting up your organization, teams, invites, and provider connections. Only the first step is required; you can do the rest later. - Connect your AI provider accounts
To read your organization's billed AI spend, Keld Atlas needs a read-only admin API key for each provider. Keld supports Anthropic and OpenAI. Google Gemini spend is estimated from telemetry and needs no key, and AWS Bedrock is coming soon.
Important: Keld's access is read-only. Keld cannot see your prompts or send prompts on your behalf through the integration connection.
For step-by-step instructions, see Integrations in the Keld Atlas Administrator Guide.
Expected result: the connection appears in Integrations with a recent sync timestamp and no warning icon. - Invite your team
Click Teams in the sidebar to invite people by email and assign roles, or import members in bulk from a CSV file during onboarding. Each person receives a join link by email. See Teams in the Keld Atlas Administrator Guide.
Expected result: pending invitations are listed on the Teams page. - Have your team install Keld Signal
Provider connections alone show what you spent, but not who spent it. Attribution by person, team, and repository requires Keld Signal on each user's computer. Send your team the next section.
Connect your computer
Every person whose AI usage you want to measure completes these steps on their own machine. To send your AI usage to your organization's Keld Atlas dashboard, follow these steps:
- Accept your invitation
Open the join link from your invitation email and sign in to atlas.keld.co. - Download Keld Signal
In the Keld Atlas dashboard, click Connect tools and download the installer for your platform. - Run the installer
Sign in with your Keld Atlas identity when prompted. The installer detects the supported tools you have installed and configures each one to send telemetry automatically. You do not need to edit configuration files or restart your tools manually.
For more detail, see Installing Keld Signal. - Verify that your usage is arriving
Run one prompt in a supported tool, then open the Activity view in Keld Atlas. Your prompt should appear within seconds, attributed to your name.
Expected result: a new entry in the Activity feed with your identity, tool, and model. If nothing appears, see Troubleshooting below.
Analyze your spend
Once at least one provider connection or Keld Signal installation is reporting, Keld Atlas begins analyzing your spend and displaying it in your dashboard. Provider connection data syncs hourly, and you can trigger a manual sync at any time from Integrations, which also shows the last sync timestamp for each connection. Tool telemetry appears in near real time as prompts complete.
Important: during the preview, Keld does not import historical spend. Your dashboard starts from the moment you connect a provider or install Keld Signal, so expect the Briefing and Observability views to be sparse at first and to become more useful over the first few days of use. The Activity feed is populated as soon as someone runs a prompt.
Start with the Briefing view for a finance-first summary, use Activity to watch usage as it happens, and use Observability to see how spend breaks down across Dimensions. You can also export data from Keld Atlas as CSV for use in other analysis tools.
Troubleshooting
No data appears in the Activity feed after installing Keld Signal.
Confirm that the tool you used is supported (Claude Code, Codex, or Gemini CLI) and that you signed in during installation. In the Keld Atlas dashboard, the Connect tools panel shows a status icon indicating whether data is flowing from your machine.
A provider connection is rejected or shows a warning.
Check that you used an admin key with the correct scope, not a standard API key. Anthropic keys start with sk-ant-admin… and need usage read scope; OpenAI keys start with sk-admin-… and need usage and billing read scope. A warning on the Integrations navigation item indicates an authentication, error, or partial-sync state. Replace the key from Integrations; your existing history is preserved.
Spend appears in Keld Atlas but is not attributed to anyone.
Provider connections report billed totals without identity. Unattributed spend usually means the people generating it have not installed Keld Signal yet. Check the unattributed spend figure on the Briefing to see how much of your total is affected.
Gemini spend looks approximate.
It is. Google does not expose the billing data Keld needs, so Gemini spend is estimated from Gemini CLI telemetry rather than pulled from your account.
Coming Soon to Keld
Keld will be a marketplace where buyers and sellers meet to transact AI inference securely and efficiently. We'll add the following products to our lineup throughout 2026. Look for preview releases soon.
Keld Relay
Keld Relay is the prompt routing layer that connects your inference workloads to the Keld Exchange. Keld Relay handles order creation, price discovery, prompt routing/result delivery, and provider dispatch. Keld Relay will be available as software SDKs for all major programming languages and also in the Keld Atlas dashboard.
Keld Exchange
Keld Exchange is the core marketplace: a multi-dimensional central limit order book (CLOB) that matches AI inference demand (buy orders from organizations) against supply (sell orders from model providers). Keld Exchange is strictly neutral; orders match on price and time priority.
Keld Trade
Keld Trade is the sell-side product (SDKs and dashboard) for model providers. It gives providers the tools to list their inference capacity on the Exchange, manage sell orders, and gate execution based on real cluster utilization. Microbatching lets providers without a native batch endpoint participate in the batch marketplace as first-class sellers.