/lab
← ls ./projects

periscope-mcp

● live

Web-app QA, testing & analysis for AI agents — 74 Playwright tools that give hard verdicts, not screenshots to squint at.

  • Python
  • AGPL-3.0
  • MCP
  • QA
  • E2E
  • Playwright
  • accessibility
  • web-vitals
repository ↗

periscope-mcp is a Model Context Protocol server that gives an AI agent 74 Playwright tools to QA, test, and analyze web apps — static sites, SPAs, and apps behind a login — with headless Chrome underneath.

The point isn’t “browser control over MCP” — playwright-mcp already does that. periscope’s tools are shaped for an agent consumer wearing a QA hat: they return structured verdicts and honest errors instead of raw page state the agent has to interpret. It encodes the testing knowledge an agent would otherwise reinvent every session.

What makes it different

  • Hard verdicts, not screenshot-squinting. assert_condition returns passed: true/false plus the actual value; the checks return structured issues, not pixels to judge.
  • One call instead of ten. auto_fill_form detects fields, infers realistic test data, and fills a whole form in one call; interact_and_test batches 25 action types with checks; test_project crawls and audits a whole site.
  • Real web-app testing. Persistent authenticated sessions (form / basic / cookie), plus a visible interactive login for the flows you can’t script — 2FA, SSO, CAPTCHA — that then runs headless. Multi-step flows, network mocking, and state snapshots included.
  • Honest responses. A drag that silently did nothing, an expired session, a page that never settles — they come back flagged, never as fake success. Errors name the cause and the fix.
  • Real INP. Because it drives real interactions, it measures genuine Interaction-to-Next-Paint from Event Timing entries — something Lighthouse can’t do in lab mode (it falls back to TBT). Export the per-interaction time series as JSON or CSV.
  • Audits a browser binding can’t give you. Accessibility (ARIA validity, keyboard nav, WCAG contrast), SEO (OG/Twitter, JSON-LD, robots.txt, site-wide duplicates), GEO / agentic-search readiness (AI-crawler access, llms.txt, WebMCP), plus real Google Lighthouse.

Battle-tested by an AI agent

Every feature was hardened by pointing a second AI agent at the server and having it test real sites. Each bug it hit doing real work became a GitHub issue — all fixed with regression tests, backed by a real-browser CI suite. That loop caught a whole class of “the tool reported success but nothing happened” bugs that unit tests never would.

Get it

Open-source under AGPL-3.0 → github.com/segentic-lab/periscope-mcp. Works with any MCP client — Claude Code, Cursor, Codex, or your own agent.

One line clones and installs (auto on Debian/Ubuntu, guided elsewhere), writing your MCP client config with the right paths:

git clone https://github.com/segentic-lab/periscope-mcp && cd periscope-mcp && ./install.sh

It ships with an AGENTS.md you paste into your agent’s system prompt — the session model, the debugging playbook, and the ordering rules that trip agents up. If it’s useful, a ⭐ on the repo helps.