Jira-Playwright AI Agent
AlexandrAQA/jira-playwright-ai-agent
Converte tickets Jira em testes E2E verdes automaticamente
Ver no GitHubMCP server que transforma tickets Jira em testes Playwright E2E automatizados. O agente lê um ticket, explora a aplicação web com um navegador real, gera e executa testes, corrige até ficarem verdes e escreve o resultado de volta no Jira. Ideal para automação de QA sem intervenção manual.
Instalação
npx convez-ai AlexandrAQA/jira-playwright-ai-agent1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Automação de testes E2E a partir de requisitos em Jira
- Redução de trabalho manual de QA
- CI/CD com geração automática de testes de aceitação
Veredito

Solução inovadora que aproveita bem o MCP do Playwright e a capacidade de código do Claude para fechar um loop completo: Jira → teste → resultado. Requer setup robusto (credenciais Jira, ambiente Playwright) e funciona melhor em projetos com fluxo de tickets bem estruturado.
README
A tradução para o português deste README ainda está na fila.
AI QA Agent: Jira -> Playwright -> Jira
An AI agent that turns a Jira ticket into a passing end-to-end test, hands-free. It reads a ticket, explores the web app through a real browser, generates and runs a Playwright test, makes it green, then writes the result back to Jira and moves the ticket across the board. A human stays in the loop for irreversible actions.
Target app under test: SauceDemo.
Demo
A short screen capture of a full run (Jira ticket -> generated test -> green -> ticket closed) will be added here.
Architecture
| Role | Component | Responsibility |
|---|---|---|
| Brain + loop | Claude Code | Decides, writes test code, runs the workflow |
| Instructions | CLAUDE.md | The playbook: workflow steps and rules |
| Hands + eyes | Playwright MCP | Drives a real browser, inspects the DOM, finds selectors |
| Door to Jira | src/jira.ts | Thin REST v3 client: read / move / append |
| Secrets | .env | Jira credentials and SauceDemo test logins |
MCP (Model Context Protocol) is an open standard for connecting tools and data to a model. Playwright MCP is what makes the agent reliable: instead of guessing selectors from the ticket text, it opens the real page and picks real selectors.
Workflow ("pick up aiqa N")
- Read the ticket from Jira (
src/jira.ts get). - Move it to In Progress.
- Explore SauceDemo via Playwright MCP and find real selectors.
- Generate
tests/generated/aiqa-N.spec.ts(each ticket step is atest.step). - Run with Playwright and fix until green.
- Append the result to the ticket description.
- Move it to Done.
Tech stack
- Playwright + TypeScript (Chromium, html + json reporters)
- Playwright MCP (browser control)
- axios + dotenv (Jira REST API v3 client)
- Claude Code as the agent runtime
Setup
npm install
npx playwright install chromium
cp .env.example .env # then fill in your Jira values
The Playwright MCP server is already declared in .mcp.json (it points to
playwright-mcp.config.json for the browser launch flags), so Claude Code picks it up
automatically when you open this folder.
Required .env values: JIRA_BASE_URL, JIRA_EMAIL, JIRA_API_TOKEN, JIRA_PROJECT_KEY.
Create the API token at id.atlassian.com -> Security -> Create API token.
Usage
# Jira helper (CLI)
npx tsx src/jira.ts get AIQA-1
npx tsx src/jira.ts move AIQA-1 "In Progress"
npx tsx src/jira.ts append AIQA-1 "Automated and passing."
# Run tests
npm test # all tests
npm run report # open the HTML report
Then, inside Claude Code in this folder, just say pick up aiqa 1.
Label-triggered autonomous mode
Instead of running each ticket by hand, a polling watcher picks up tickets automatically.
Add the label playwright_agent to a Jira ticket and the agent runs the whole flow on its own.
# add the trigger label to a ticket
npx tsx src/jira.ts label-add AIQA-2 playwright_agent
# safe dry run (detects labeled tickets, prints the command, runs nothing)
WATCH_DRY_RUN=1 WATCH_ONESHOT=1 npx tsx scripts/watch-jira.ts
# start the watcher (polls Jira, runs the agent on each new labeled ticket)
npx tsx scripts/watch-jira.ts
scripts/watch-jira.ts polls Jira with
project = AIQA AND labels = playwright_agent AND status = "To Do" and, for each new ticket,
runs the agent headless:
claude -p "pick up AIQA-N ..." --model haiku --permission-mode bypassPermissions \
--mcp-config .mcp.json --strict-mcp-config
Configurable via env: WATCH_MODEL (haiku | sonnet | opus), WATCH_INTERVAL_MS, WATCH_LABEL.
This is the polling variant of the trigger; a Jira webhook would be the event-driven alternative.
Project structure
.mcp.json Playwright MCP server definition
playwright-mcp.config.json browser launch flags for the MCP server
.env / .env.example secrets (env) and template
playwright.config.ts Playwright config (chromium, reporters, baseURL)
CLAUDE.md agent playbook (read automatically by Claude Code)
src/jira.ts Jira REST v3 helper + CLI
scripts/seed-tickets.ts one-off: seed ticket descriptions
scripts/watch-jira.ts polling trigger (label-based autonomous runs)
tests/generated/ generated Playwright specs
Notes
This is the lightweight variant: Claude Code as the agent, a hand-written Jira REST client, manual trigger. It evolves naturally toward a production setup (custom agent loop on the Anthropic SDK, Jira via MCP, label-based trigger, CI, human-in-the-loop).