Pre-deploy Checks (MCP)

You ship fast and you don't have a test suite. Pre-deploy launch checks let your coding agent (Claude Code, Cursor, Windsurf) make sure your critical flows still work before you hit deploy — so you ship without holding your breath. You connect the Opsis MCP server (opsis-mcp) to your agent; the agent reads your codebase and derives the flows that matter — checkout, signup, your app's primary action — from real routes, forms, and selectors. Opsis then runs each flow in a real Chromium browser on your machine and captures full evidence: video, trace, screenshots, console and network logs.

How a launch check flows

1. Your agent calls opsis_launch_check_guide and derives 2–6 critical flows from the codebase — routes, forms, and the real selectors in your JSX/HTML, not guesses from URLs.
2. For each flow it calls opsis_run_flow_check. Playwright runs the flow locally against your dev or preview server, recording video, trace, screenshots, console, and network evidence.
3. A flow passes when every step succeeds with no console errors and no failed requests. On failure, the agent gets the evidence paths plus a guard-railed safe-fix prompt — it applies the fix and re-runs the same check.
4. When all flows have run, the agent submits the aggregate report with opsis_submit_launch_report. The Pre-deploy checks stage on your project dashboard updates instantly with the verdict: Launch-ready: PASS or N ISSUES FOUND.

Setup (once per project)

Pre-deploy checks belong to a project, so create the project first by connecting PostHog with OAuth. Then:

1. Open your project's Before Ship tab and expand “Run the check on your machine with your coding agent”, then reveal the MCP key (opk_…). The key authorizes only this project's launch-report API — never your PostHog data or the dashboard.
2. Add the server to your coding agent. Both snippets on the Before Ship tab are pre-filled with your project's values, so you can copy them as-is.

   Claude Code (one command):

   claude mcp add opsis \
     -e OPSIS_APP_URL=https://<your-opsis-domain> \
     -e OPSIS_PROJECT_ID=prj_xxxxxxxx \
     -e OPSIS_MCP_KEY=opk_xxxxxxxx \
     -- npx -y opsis-mcp

   Cursor / Windsurf (mcp.json):

   {
     "mcpServers": {
       "opsis": {
         "command": "npx",
         "args": ["-y", "opsis-mcp"],
         "env": {
           "OPSIS_APP_URL": "https://<your-opsis-domain>",
           "OPSIS_PROJECT_ID": "prj_xxxxxxxx",
           "OPSIS_MCP_KEY": "opk_xxxxxxxx"
         }
       }
     }
   }

3. Make sure Chromium is installed on your machine:

   npx playwright install chromium

4. Start your dev or preview server, then ask your agent, e.g.:

   Run Opsis launch checks against my dev server on http://localhost:3000.

Evidence is stored locally under .opsis/predeploy/ in the repo your agent runs from. Check the latest result anytime with opsis_launch_status — or on the project dashboard, where the Pre-deploy checks stage shows Passed / Failing / Flaky with flow counts.

The launch-check tools

These four tools drive the launch-check loop above. The same MCP server also exposes a separate Pre-deploy Security Scan (an OWASP Top 10 check), documented on its own page.

What pass and fail mean

• Pass: every step in the flow succeeded, with no console errors and no failed network requests.
• Fail: a step failed, a console error appeared, or a request failed — the agent receives local evidence paths and a safe-fix prompt scoped to what the evidence shows.
• Verdict: Launch-ready: PASS when every flow passes; N ISSUES FOUND lists how many flows need attention. The dashboard stage mirrors this as Passed, Failing, or Flaky.

Privacy

• The Opsis MCP server never reads your codebase — your coding agent does, on your machine.
• Only redacted flow results are sent to Opsis: title, route, status, a one-line summary, plus redacted console/network/page error lines and a copy-paste fix prompt. The raw video, traces, screenshots, and full logs stay local under .opsis/predeploy/.
• fill values must be synthetic (the guide enforces this) — fake emails like test@example.com, never real user data.
• The opk_ key is scoped to one project's launch-report API. It cannot read PostHog data, findings, or anything else in your dashboard.

Troubleshooting