Skip to main content
A coding agent with the Momentic MCP server and the momentic-spec skill turns a user story into a .test.yaml file before anyone writes product code. The agent reads the story, maps the acceptance criteria to steps and assertions, and writes the file into your repository. Once the feature is usable in your app, the agent prepares the test state and runs the smallest affected set at an end-to-end checkpoint. Long runs require confirmation. The test is plain YAML, so you edit it in a pull request like any other file.

​
Setup

Install the MCP server and the skills with the CLI. Both commands are described in Building with AI:
The agent needs a MOMENTIC_API_KEY in its environment and a reachable app to test: a local dev server, a preview deploy, or staging.

​
In your coding agent

Inside Claude Code, Cursor, Codex, or another agent with the Momentic MCP server and skills installed, paste the story after the skill name:
The agent updates existing coverage or creates the smallest test set that expresses the changed behavior. For a feature that is not implemented yet, it marks the specification disabled until its checkpoint is ready:
apply-promo-code.test.yaml
The second entry uses fill, which replaces the field’s current value instead of appending to it. See type. The agent does not run the test right away. Momentic runs are end-to-end checks against the real UI, so the skill waits until the feature is implemented and a user can exercise it in the app. At that point the agent prepares the required data and account state, enables the affected tests, and proposes the smallest relevant run. It asks for confirmation before a long run; an explicit request to run those tests counts as confirmation.

​
Edit the generated test

The file lives in your repository, so review it in the same pull request as the feature:
  • Rename the id to something your team recognizes, then keep it stable. See File format.
  • Replace a natural-language step with a preset step when the action is exact. click: Apply gives the runtime one action to perform, where “click the apply button” leaves the path to an AI action. Steps lists the preset steps.
  • Match each check to the contract in the story. If the story fixes an exact total, verify it with an element-content or JavaScript check rather than a qualitative AI assertion.
  • Move shared setup such as log-in into a module so every test generated from a story reuses it.
Once the feature is ready, remove disabled: true and run the edited file locally before you push:

​
What to expect

  • Give the agent explicit acceptance criteria. Clarify missing expected behavior before treating a generated test as the product contract.
  • Both AI actions and preset steps can target elements by description. Keep a future specification disabled until the required UI and dependencies exist.
  • When the UI changes later, locator auto-healing re-resolves a stale target during the run, and failure recovery can clear an obstruction and retry. See AI test maintenance.
For tests scoped to a code change instead of a story, see Generate tests from a pull request diff.