The explore agent (
momentic ai explore) is deprecated. Mo now handles
autonomous bug finding. Sign up for access. To write
tests from a code diff, use the MCP server with the momentic-spec
skill.install-mcp registers two servers, including in Claude Desktop. Agents use
both:
Prerequisites
For the local server:- Momentic CLI installed locally
- A project with a valid
momentic.config.yaml - A
MOMENTIC_API_KEYor an existing Momentic CLI sign-in available to the MCP process. Configure API keys in that process’s environment rather than relying on the editor to inherit your shell environment.
npx -y qa mcp with the shared CLI login. It does not
require a project config. Install Momentic, run npx -y qa login, then add MCP.
qa login, momentic login and momentic-mobile login share the saved login;
these steps use npx -y qa login. Claude.ai/ChatGPT connectors can instead use
https://mcp.momentic.ai with OAuth.
The Momentic plugin also keeps the
hosted URL and OAuth by default. Plugin users can
opt into the CLI bridge
after disabling the plugin’s hosted entry. Do not run install-mcp alongside
the plugin; it duplicates the local servers.
Setup instructions
Runinstall-mcp to detect the coding agents on your machine. It writes the
local and hosted server entries into each one’s config, including Claude
Desktop. Use momentic-mobile install-mcp for mobile projects; it registers a
separate momentic-mobile server, so web and mobile can live in the same
config.
Both commands share momentic-hosted for run history, quarantine, and Mo. New
hosted entries run npx -y qa mcp with your saved CLI login; no API key is
written to this entry.
Existing hosted URL entries switch to the bridge only when a saved login has
both WorkOS access and refresh tokens. Without those tokens, the installer keeps
the URL entry and its OAuth sign-in. To switch, run npx -y qa login, then
rerun install-mcp with the same options.
Existing hosted connections keep their names unless named momentic or
momentic-mobile. These reserved slots are not migrated: the selected one
becomes local, and the other is untouched. Back up their hosted settings before
installing; a new momentic-hosted entry uses defaults.
momentic install-mcp for
--client, --scope, and noninteractive use. The tabs below show the command
for each client and the manual config it writes.
The local entry uses stdio. Clients launch it with one of these commands:
Web
Mobile
momentic
with momentic-mobile in the command and use momentic-mobile as the server
name.
Keep the shared momentic-hosted entry unchanged for mobile projects.
See momentic mcp and
momentic-mobile mcp for every
supported flag (headful browser, pixel ratio, daemon mode, etc.).
Both servers accept --headful false. For web, this launches headless browser
sessions. For mobile, it skips the remote-control viewer without changing the
emulator UI. Override the server default with headful when starting a session.
Older releases use the names listed in CLI help and the tool schema:
--headful-browser for web and headfulBrowser for session overrides. Those
names remain supported.
After registering the servers, verify each connection:
- Restart the client and start a new chat.
- Ask the agent to call
momentic_get_artifactson the localmomenticserver (web) ormomentic-mobileserver (mobile). Check the returned project artifacts. - Separately call
momentic_search_docsonmomentic-hosted.
- Claude Code
- Cursor
- VS Code
- Codex
- OpenCode
- Devin
- Other
claude mcp list.Project servers can show “Pending approval”. Open /mcp in Claude Code and
approve momentic (or momentic-mobile) and momentic-hosted.On native Windows, wrap the local stdio launcher with cmd /c:MOMENTIC_API_KEY via the CLI’s --env flag or by running the command
from a shell that already has it exported. See
Claude Code MCP docs
for scoping and auth options. Run /mcp to confirm the server is active.Example prompts
Choose a task and copy its prompt into your coding agent. Replace the test filenames and run ID with your own. For longer jobs, invoke a skill by name. The skill tells the agent which tools to call and in what order.Check the MCP connection
Preview a test
Run and repair a test
Inspect a past run
Lint test files
Look up documentation
Write your first test with the agent
The agent writes a.test.yaml, runs it through the MCP server, reads the
structured failure with the failed step, screenshot, and trace, then fixes its
own test.
Verify the server with claude mcp list in Claude Code, or turn it on in
Cursor’s MCP settings. Start a new chat. This web example uses an email-only
login flow. Replace the URL and test account with your app’s values, and adapt
the steps to your login flow.
1
Write the test
Ask the agent to write a test with a base URL:
2
Run and repair the test
Ask the agent to run the test and fix any failures:
3
Commit the test
Review the resulting file, then commit it:
momentic-test skill covers the
authoring flow, and GitHub Actions runs the
test in CI. The test opens the top-level url before it runs the steps:
login.test.yaml
Tool reference
Use the localmomentic or momentic-mobile server for repository tests. Both
means web and mobile; no registered tool name is mobile-only.
For hosted tools, see the
hosted MCP tool reference.
Usage tips
Choose a model that can call tools and follow the installed skills. Reasoning settings affect latency and decisions; compare them on your own tests before changing the team default. Invoke the relevant skill by name when authoring, maintaining, or changing product behavior. Be specific in prompts, and include a step limit for exploratory runs. Let the agent close each session’s browser; hard-stopping your editor leaks browser processes. Add.momentic-mcp/ to .gitignore. MCP stores screenshots, videos, and traces
there.
Tell the agent to call momentic_search_docs on the hosted Momentic MCP (via
qa mcp) before it writes step options, CLI flags, or configuration keys from
memory. The tool searches these docs and returns the matching sections with
their links.
Step previews and step execution wait 30 seconds by default, which you can
change with timeoutSeconds. Work that takes longer continues in the background
and the tool returns a stepRunnerId.
Poll for the result, and pass the polling request’s timeoutSeconds (0-30,
default 0) to hold the poll open instead of busy-polling. Only one preview or
run executes per session at a time.
Use the hosted Momentic MCP (via qa mcp) to list runs and quarantines. For
mobile run history, pass executionType: ["ANDROID", "IOS"] or select one
platform explicitly. Both local servers let you download and inspect past runs,
their steps, and step results on your machine.