By default, Docker containers are on their own network and access the host via host.docker.internal.

The existing (and non-obvious) workaround is to replace the Access URL in the init_script. See below:

  command = [
    "sh", "-c",
    <<EOT
    trap '[ $? -ne 0 ] && echo === Agent script exited with non-zero code. Sleeping infinitely to preserve logs... && sleep infinity' EXIT
    ${replace(coder_agent.main.init_script, "localhost", "host.docker.internal")}
    EOT
  ]

This approach is brittle because it strongly couples the template with the Access URL. In this example, if the Access URL is 127.0.0.1 and not localhost, the agent will never start.

@kylecarbs suggested we add a host attribute the coder_provider block that would contain the host.docker.internal isntead.