Skip to content
Debug failures

Fix in your agent

Four ways to get AI help with failing tests, when to use which, and a full flow with the MCP server.

Route Use it when Plans
Fix with Squally in the attempt page’s Fix panel you want a suggested fix on the attempt page, without leaving Squally Standard, Pro
Fix in your agent in the attempt page’s Fix panel you want your coding agent to fix it, with your repository open: With MCP (the agent loads the failure itself) or Without MCP (the whole prompt), then Copy prompt every plan
Fix in your agent in the run page’s Fix panel your agent should fix all of a run’s failed and flaky tests, each distinct error once: one prompt through the MCP server Standard, Pro
MCP server your agent should look things up in Squally itself - which run, which test, how often Standard, Pro

The Fix panels start from a page in Squally. The MCP server starts from your agent: you describe the problem, the agent asks Squally.

With the MCP server set up in your agent, ask in plain words - for example “Why did the last CI run on main fail?”. The agent works through Squally’s tools:

  1. squally-list-projects - the project and its ID.
  2. squally-find-run - the run, by branch, commit or status.
  3. squally-get-run - the run’s failing tests.
  4. squally-debug-failure - every attempt of one failing test, with its errors, stacks and the same prompt as Without MCP in the Fix panel.
  5. With your repository open, the agent proposes a fix; you review and push.
  6. After the next runs, ask again: squally-get-test-metrics shows the test’s stability, flaky rate and failure rate - stability should go up.
  • The latest squally-mcp. Each version is made for one read API version, and app.squally.dev always serves the newest: an MCP client that checks the tools’ structured output - the MCP SDK’s does - rejects the answers of an older package. npx -y squally-mcp fetches the latest; a global install keeps its version until you run npm install -g squally-mcp again. Which version needs which API: the changelog. It needs Node 22 or newer.
  • An organization read key (sqly_ro_…) on the Standard or Pro plan. The server refuses a project ingest key. See Keys.