The CLI
theformalera command and the Python annotation SDK (import theformalera as tfe). It needs Python 3.10 or newer and runs on Linux and macOS.
A CI runner that sends work to the cloud solver needs only
theformalera[triton]. A runner that solves locally needs theformalera[triton,engine].
Annotations are no-ops at runtime. Your inference code can import
theformalera in production without pulling in Triton, Lean or the engine.Lean 4
theformalera pull builds each spec package as a Lean 4 (Lake) workspace, and local verification checks every proof with the Lean kernel. Install Lean with elan:
lake on your PATH and then in ~/.elan/bin. The toolchain version comes from each spec package (leanprover/lean4:v4.35.0 today), and elan installs it on first use.
You can skip Lean when you use --runner cloud: proofs are checked in the cloud, and theformalera pull --no-build skips the local build. You still need Lean to re-check proofs yourself.
Log in to Spec Hub
~/.config/theformalera/credentials.json (mode 0600). In CI, set THEFORMALERA_TOKEN instead. Public specs and the public lemma library can be read without a token; cloud verification and the org lemma library need one.
Check the install
engine extra, the second part reads theformalera engine (remote).
