Skip to content

qawolf/qawolf

v0.1.14

Ask QA Wolf for test coverage, run flows, read results, and drive a cloud browser from your editor.

GitHubOpen Website

QA Wolf plugin

Use QA Wolf from your coding agent to request test coverage, run flows, inspect results, and drive a cloud browser. See platform setup for native plugins, MCP configuration, and guidance-only limitations.

Install it from the QA Wolf plugin repository through your client's plugin marketplace. The plugin signs in with OAuth.

Connect to QA Wolf

QA Wolf clients connect to https://app.qawolf.com/api/mcp and sign in with OAuth through https://signin.qawolf.com. Codex offers the sign-in inside the conversation the first time it calls a QA Wolf tool. The ChatGPT app, Claude Code, Claude Desktop, and Gemini ask when the server is added or on first connection.

No API key is needed. Do not add an Authorization header to the plugin's MCP entry; in Claude Code and Codex a configured header switches OAuth off and the connection fails with HTTP 401.

Where a browser sign-in cannot happen, such as CI or a container, use the API key fallback described in platform setup. Keep any key out of shell history, source control, chats, issues, and screenshots.

OAuth reaches every workspace you are a member of, across all of your organizations. whoami lists them with the organization that owns each one. When there is exactly one the connection binds to it; otherwise pass workspaceId, and a browser tool binds to the workspace you name. A team API key is always bound to its team.

Install in Claude Code

/plugin marketplace add qawolf/agent-plugins
/plugin install qawolf@qawolf

Start a new session after installation. Open /mcp, select the QA Wolf plugin server, and choose Authenticate if sign-in is needed. Follow the browser sign-in, then return to Claude and ask it to call whoami to confirm your account and workspace.

Claude Code handles the callback and token storage. If the browser does not open, use the link in its authentication UI. If the redirect fails, use its dedicated authentication prompt. Do not paste callback URLs, authorization codes, or tokens into chat.

To sign in from a terminal instead, run claude mcp login plugin:qawolf:qawolf; the plugin-scoped name is required, and the bare qawolf does not resolve.

Install in Codex

codex plugin marketplace add qawolf/agent-plugins
codex plugin add qawolf@qawolf

Start a new Codex session, then ask Codex to call whoami. It offers the sign-in inside the conversation on that first tool call; codex mcp login qawolf starts it by hand.

Install in the ChatGPT app

ChatGPT does not take MCP servers from an installed plugin, so add QA Wolf as a connector once. Turn on developer mode in settings, then add an MCP server with the URL https://app.qawolf.com/api/mcp and no header. ChatGPT authorizes it during setup. Then ask QA Wolf for something in a new chat.

Install in Claude Desktop

Add QA Wolf as a custom connector with the URL https://app.qawolf.com/api/mcp and no header, then authorize it.

Install in Google Antigravity CLI

agy plugin install https://github.com/qawolf/agent-plugins

Then run /mcp auth qawolf to sign in. Ask the agent to call whoami to confirm your account and workspace.

Other coding agents

Any client with MCP support connects to https://app.qawolf.com/api/mcp. The server supplies the skills, so no skill files are installed. Use the platform guide for your client, then verify the connection with whoami.

Skills

The MCP server serves the skills through two tools. skill_list names them with the description that says when to use one. skill_get reads the current version of one skill, so a skill change reaches every session without a plugin update.

  • Flow Outline (qawolf-flow-outline) handles every request to create a flow, including finishing a draft and covering a pull request. Your coding agent explores through runner computer use, finds context independently, and uses the ask-user tool for gaps. It presents Arrange, Act, Assert outlines for approval both as a message and through propose_outline, which draws them in a card in the clients that support one, sends approved outlines through agent_send, and monitors creation.
  • Flow Maintenance (qawolf-flow-maintenance) repairs a flow that already exists and has started failing. It reads the recorded run and diagnosis, then hands the fix to agent_send without launching a runner.
  • Onboarding (qawolf-onboarding) chooses the best first flow and reads Flow Outline, then hands off to Trigger Setup once the flow is active. A request that already names the new flow goes directly to Flow Outline.
  • Trigger Setup (qawolf-trigger-setup) makes flows run automatically. It reads the repository's CI, recommends a deployment trigger through GitHub or GitLab Deployments or a deployment.reportStatus call, and finishes with a scheduled trigger when the pipeline will not change.
  • Trigger Diagnostics (qawolf-trigger-diagnostics) answers why a trigger did not run. It reads the deployments QA Wolf received and each trigger's recorded verdict, explains the outcome in the user's terms, and changes nothing; the fix it proposes is made elsewhere.
  • Trigger Migration (qawolf-trigger-migration) moves a workspace off legacy triggers. It reads every legacy trigger, presents one complete plan as a dry run, creates every replacement only after approval of the whole plan, and pauses the legacy triggers afterwards without deleting anything.
  • QA Wolf (qawolf) supplies shared connection and safety guidance, plus existing-test operations.

Your new-flow or onboarding request covers billed browser use and routine staging exploration without a separate permission prompt. Your coding agent displays the full AAA before asking for approval, then shares the session URL before monitoring. It continues silently when status is unchanged and verifies publication and readiness before reporting completion. An onboarding first flow becomes active once its run passes, with no draft-or-active question; a flow you ask for by name still lets you choose. Activation makes a flow eligible for triggered suites, so it runs on its own only once Trigger Setup or an existing trigger covers it. Move any flow back to draft in the app.

Try it

  • "Onboard this app to QA Wolf and find the best first flow."
  • "Create a QA Wolf flow for checkout."
  • "Run the smoke-tagged QA Wolf flows and summarize what failed."
  • "Open a QA Wolf browser and reproduce the login bug."

A cloud browser bills while its runner exists. The skill instructs the agent to terminate the runner when the work ends.

Troubleshooting

  • Connection failure or HTTP 404: check the configured URL. Reinstalling the plugin cannot fix an unavailable service.
  • Sign-in never starts and the connection reports HTTP 401: an Authorization header is configured somewhere. Claude Code and Codex skip OAuth when one is set. Remove it from your own qawolf MCP entry and reconnect.
  • Sign-in fails or the session expires: run claude mcp login plugin:qawolf:qawolf, or codex mcp login qawolf, and complete the browser flow again.
  • Browser-tool authorization error: the connection reaches several workspaces, so it is not bound to one. Call whoami, then pass workspaceId for the workspace you want, or use a team API key already bound to it.
  • Missing agent tools: coverage requests require both agent_send and agent_get. Contact QA Wolf support if either is unavailable.

Use the platform guide for client-specific connection settings. Never send a QA Wolf API key or OAuth token to an untrusted endpoint.

Report plugin problems through GitHub issues. Do not include API keys, passwords, or customer test data in public reports.

Privacy and terms

The plugin sends tool arguments to QA Wolf and returns the requested API results to your coding agent. Depending on the tools used, this can include test data, environment variables, and browser screenshots. Review the requested actions before approving them.