AI in integrator.io is metered three ways: AI agent actions count what an agent did, AI credits measure the AI processing it used, and, on plans that include MCP, MCP calls count what the MCP servers you publish receive. This article explains what consumes each, why the same agent can use different amounts of credits on different runs, what happens when you reach a limit, and how to keep credit use predictable. Go to Account , and then select Subscription to view your AI actions used, AI credits remaining, and MCP calls.
AI agent actions
An action is one tool execution by an agent during a flow run:
- LLM native tools – Tools the model provides, such as image generation or web search.
- MCP tools and resources – Calls to tools, or retrieval of resources, through an MCP server.
- Celigo tools – Tools added directly to the agent.
- Retries and loops – Each repeated attempt after a failure is another action.
Actions used while testing or previewing an agent, or in non-production environments, aren't counted.
Actions count what a Celigo AI agent executes. When an agent invokes a tool on an MCP server published in your account, the execution is an action for the agent and, separately, an MCP call for the server (see MCP calls).
The actions entitlement is a soft limit: your agents keep running above it, and you're notified by email as you near, reach, and exceed it.
Review your AI actions
On the Subscription page, select View next to AI actions used. Current usage lists this month's actions by agent, with the Name of the agent, the Provider of the tool (Celigo for Celigo tools, External for MCP tools), the Type (Agent, Tool, or Resource), and the Capability name of the tool; filter by provider or type to narrow the list. Historical usage shows actions by month.
AI credits
Credits are consumed when an agent does AI work during a flow run: model reasoning, tool calls, and response generation. Creating or editing an agent and saving its configuration don't use credits, but a test that triggers AI does.
A run uses more credits when it includes more AI steps (several reasoning or generation calls), larger inputs (long prompts, big payloads, logs, or documents), longer outputs, or retries after failures. Because of this, two runs of the same agent can cost different amounts: the input data differs in size or complexity, the agent takes a different path, a tool failure triggers retries, or the prompt produces a longer answer.
Caution: AI credits are an enforced balance. When it reaches zero, agents stop running and the error insufficient AI agent prepaid credits is shown. You're notified by email as you near, reach, and exhaust the balance. To add credits, select Request upgrade on the Subscription page or contact your account team.
Review your AI credits
On the Subscription page, select View next to AI credits remaining. Current usage lists this month's consumption by agent, with the Name of the agent, the Environment it ran in (filterable), and the AI credits consumed; the same agent appears once per run, so you can compare runs. Historical usage shows credits by month.
MCP calls
If your plan includes MCP calls, your subscription allows a set number of calls per month to the MCP servers published in your account, pooled across all of them; the MCP calls row appears on the Subscription page only on those plans.
- Each tool, prompt, or resource invocation received by one of your MCP servers counts as one call, whoever makes it: an external client or a Celigo AI agent.
- Failed invocations count. Retrying a failed call adds another call.
- Calls are pooled across every MCP server in the account; there's no separate allowance per server or per connected user.
MCP calls and AI agent actions are separate meters that can both apply to one call. When a Celigo AI agent invokes a tool on your MCP server, the server records an MCP call and the agent records an action. A call from an external client is an MCP call only; if it triggers a Celigo agent, the actions that agent performs are counted as actions. A Tool called through an MCP server doesn't count the connection it uses as an endpoint (see Track your endpoint usage).
The MCP call entitlement is a soft limit: you're notified as you approach, reach, and exceed it, but calls aren't blocked. Rate limits and usage quotas that you configure on an MCP server are separate hard limits: when a server's limit is reached, it rejects calls with HTTP 429 regardless of your entitlement. See Manage MCP servers.
Keep credit use predictable
- Set limits where the agent allows them, such as a maximum number of tokens per run.
- Keep prompts and context to what the task needs; don't pass large logs or datasets unless the agent has to read them.
- Prefer structured inputs over free text.
- Ask for concise output formats, such as bullets, JSON, or a short summary, unless a detailed response is required.
To estimate a run's cost, test the agent on representative input and check the credits consumed. Runs on input of similar size and complexity cost about the same. If runs cost more than expected, look for repeated model calls in a loop, large payloads in the context, prompts that ask for long answers, or retries caused by tool or validation failures.
Frequently asked questions
Are AI credits the same as agent actions? No. Actions count tool executions (what happened); credits measure AI processing (what it cost). MCP calls are a third meter, for calls received by your MCP servers (see MCP calls).
Are credits used if an agent doesn't do AI work in a run? No. Credits are consumed only when AI work is performed.