AI credits
What draws an AI credit, which workspace pays for it, what happens when a workspace runs out, and when the count resets.
This page tells you which actions draw AI credits, so you can predict what a month of work costs and understand a refusal when one arrives.
What a credit is
A credit is one AI action. Every workspace has a number of them each month, shown in Settings → Billing next to the count drawn so far. The count resets on the first day of the month, UTC, whatever day the plan renews on.
What draws a credit
| Action | Drawn when |
|---|---|
| Enrichment | A requirement is enriched, whether you created it by hand, through the API, from Slack, from a Jira, Linear or GitHub issue, or through an AI agent over MCP |
| Test generation | Test cases are generated, whether from the Generate button, from moving a requirement to In development, after enrichment, or through an AI agent over MCP |
| Import run | An import run extracts requirements from a document, a spreadsheet, pasted text or a URL, in the import wizard or through an AI agent over MCP |
| Clarification | Spectrace asks clarifying questions about text you are about to import |
| Playwright test | A Playwright test is generated for a requirement |
Spreadsheet imports draw one credit for each row enriched, so a 40-row sheet with Enhance with AI ticked draws 40. Every other action in the table draws one credit each time it runs.
These draw no credits: pull request verification, which has its own monthly allowance of checks; semantic search, which uses embeddings; the suggestions offered while you write a requirement or plan a cycle; and the progress hints the VS Code extension sends while you code.
AI agents over MCP
An AI agent connected over the Model Context Protocol draws from the same meter, at the same cost per action as the same work in the app. Extraction, ambiguity analysis and test generation each draw their credit, and a requirement an agent creates draws the enrichment credit when the gap-fill runs. A call that cannot cover its cost is refused before the model runs, with an error naming what it needed and what is left, and nothing is charged.
Inline pull request verification from an agent draws a pull request check rather than a credit. Reading requirements, toggling acceptance criteria and spectrace_semantic_search draw nothing. See Connect an agent.
Which workspace pays
Background work draws from the workspace that owns the project, whether a webhook, a scheduled job or an AI agent started it, and not from the workspace of whoever was signed in when the requirement was created.
When a workspace runs out
An action you start yourself refuses with a message naming what it needed and what is left, and nothing is charged.
Background work stops instead of running unpaid. An enrichment or test generation that cannot draw its credit leaves the requirement as it was, so the work can run later without anything to undo. The job records itself as skipped, with the reason.