01
Register the server
Momentic runs as a local MCP server. Claude Code, Codex, and Cursor register it in one line, and mobile agents use the same server.
platform
/ coding-agents
Your coding agent writes the tests and runs them.
Momentic ships a local MCP server. Cursor, Claude Code and Codex call it to open a browser, build a test step by step, save the YAML in your repo, and read the failure back.
Cursor · Claude Code · Codex · OpenCode · GitHub Copilot
github.com
tests/checkout/promo-code.test.yaml+11−0
@@ new file · written by the agent @@
+
fileType: momentic/test/v2+
id: amber-cedar-ledger+
url: http://localhost:3000/cart+
steps:+
- click: The "Checkout" button+
- type:+
text: SAVE10+
into: The promo code field+
- act: Apply the promo code+
- assert: The order total is 10% lowerclaude — ~/storefront
Agents at work in real repositories.
117,010 bugs caught before deploy
70M+
Test runs executed
8.9M+
Steps auto-healed
80k+
PRs verified
339k+
Changes verified
How an agent uses Momentic.
The agent registers one local server, reads the skills that tell it how to test, then builds the test in a real browser and leaves the YAML in your repo.
02
Install the skills
One command installs the Momentic skills. They tell the agent how to build a test, how to write the affected tests before a change, and how to investigate a failure.
03
The agent writes and runs the test
It drives a live browser, saves the YAML in your repo, runs it, and reads the result. The test change arrives in a pull request, and CI runs it on the next commit.
What the MCP server exposes.
The agent calls Momentic tools over a local stdio server, follows the skills that tell it how to test, and leaves a YAML file you review.
Tests built against a live browser
The momentic-test tools let an agent open your app, try a step, look at the result, and keep the step only when it works.
The agent sees the real page before it saves a step.
Executable specs before the code
With momentic-spec, the agent writes the affected tests before it changes user-visible behavior, so the implementation has something to satisfy.
The agent writes the check first, then makes it pass.
Failure investigation with evidence
momentic-maintain finds the runs for a branch, reads the saved classifications and run evidence, and points at the first run where behavior changed.
The agent reads the run history before it edits a test.
Web, iOS and Android
momentic-mobile mcp gives an agent the same tools against iOS simulators and Android emulators, with the same YAML format.
Same format on all three platforms.
Runs on your machine
The server is local stdio. It reads momentic.config.yaml in your project and uses the MOMENTIC_API_KEY you give the editor.
The server talks to the project on your disk.
The output is a file you review
Every test an agent writes is YAML in your repo, so it goes through code review and the CLI runs it in CI.
Agent work lands in a pull request.
Engineering teams that let agents keep the tests current.
Engineers stay in the editor, and the tests still land in the repo with the code change.
“Momentic gave us a fast and reliable way to validate Poe.com's AI responses, even when they weren't deterministic.”
Momoko F. Head of Product Operations, Quora
30 min
daily test execution, down from 7 hours
500+
manual test cases replaced
100%
critical tests created in one month
Available across every platform.
The same MCP tools and the same YAML on all three.
Coding agents, answered.
What is the Momentic MCP server?
A local stdio server that exposes Momentic as tools a coding agent can call. The agent can open a browser, create and edit tests, run them, and read run history.
Which agents work with it?
Cursor, Claude Code, Codex, Devin, OpenCode and GitHub Copilot. Each one registers a stdio server that runs npx momentic mcp.
How do I set it up in Claude Code?
Run claude mcp add --transport stdio momentic -- npx momentic mcp --config /absolute/path/to/momentic.config.yaml, restart Claude Code, then run /mcp to confirm the server is active.
What are Momentic skills?
Markdown instructions that tell the agent how to use the tools. momentic-test builds tests, momentic-spec writes the affected tests before a product change, and momentic-maintain investigates failures. Install them with npx skills add momentic-ai/skills.
Does the agent need an API key?
Yes. Most editors do not inherit your shell environment, so set MOMENTIC_API_KEY in the MCP entry or start the editor from a shell that exports it.
Where does the agent save the test?
As a YAML file in your repo, next to your code. You review it in a pull request.
Can the agent run the test it just wrote?
Yes, through the same server, in a real browser. The CLI runs that same file in CI later.
Our coding agent already runs our Playwright suite. What changes?
Your agent stops maintaining the runner. The test adapts as it runs: a Momentic locator re-resolves when an element moves, failure recovery clears a transient obstruction and retries the step, and triage classifies what broke. A moved element is handled inside the run, not by your agent. When a test does fail for a real reason, the failing step and its evidence come back over MCP, so the agent spends its turns on the fix instead of on the test. The build vs buy calculator prices the suite your agent keeps.
Is that cheaper than having the agent maintain a Playwright suite?
Yes. The run repairs itself, triages the failure, and repairs the test when it can, so a test change does not start an agent cycle. Your agent gets involved for a real product failure, and it starts at the failing step, so the loop runs less often and each cycle is shorter.
What does the agent get back when a test fails?
The failing step and its evidence: before and after screenshots, a trace with the DOM, network requests and console output, the step execution order, and the video when recording is on. The momentic-maintain skill also reads the run history for the branch, so the agent can name the first run where behavior changed.
Does the agent remember how our product works?
Momentic keeps a Knowledge base of your terminology, your rules for the agents and your known flows, and retrieves the relevant entries on every AI-assisted step. Memory adds the agent's own decisions from past runs. A new editor session starts with that context instead of reading your app from scratch, and each run makes the next one more consistent.
Does this replace the visual editor?
No. The editor writes the same YAML, so one suite can hold tests from engineers, agents and the editor.
Which CLI version do I need?
Momentic CLI 2.54.0 or later.
What is the difference between Playwright MCP and Momentic's MCP server?
Playwright MCP drives a browser and writes Playwright code, so the selectors and the strings from authoring time become the test. Momentic previews each step against the live page, keeps the step only when it works, and saves YAML in your repo that the runner replays with a step cache and auto-heal.
Can an agent run tests and read the results over MCP?
Yes. The agent runs a test through the same server, and the momentic-maintain skill reads the run history, the saved classifications and the step evidence for a branch.
Does MCP keep our tests in version control?
Yes. The agent writes a YAML file into your repo, so a test change arrives in a pull request and CI runs it on the next commit.
Can a coding agent run tests in CI over MCP?
No, and it does not need to. MCP is a local server for your editor. CI runs npx momentic run, and the exit code gates the merge.
Still have additional questions?
Contact us