Genkit and Dotprompt

The work on one slide

The work, from the public record

The contribution counts come from the repositories' author and review indexes. The implementation claims point more narrowly to the first commits, merged pull requests, release records, or current source trees that carry the work.

Dotprompt's multilingual fixed point

The implementations are not five unrelated ports. They converge on one shared file format and one executable specification: parse YAML frontmatter, render Handlebars templates, resolve partials and tools, and produce the same model request in every runtime. The spec/ fixtures are therefore the center of the system. Each language implementation is independently idiomatic, but it must agree on observable behavior.

Python could not use the usual LGPL-licensed pybars3 implementation inside an Apache-2.0 project, so the Python package binds the Apache-2.0 Rust Handlebars engine through PyO3. Dart lacked a suitable implementation, so the repository ships a pure-Dart Handlebars engine alongside Dotprompt itself. Java uses the JKnack engine; Go uses Raymond; Rust uses handlebars-rust. The format remains the invariant while the engine is an adapter.

Packages and public registry signals

Download counts are shown only where a registry exposes a comparable public figure. They are not summed across npm, PyPI, pub.dev, Maven, crates.io, and the Go module proxy: those systems count different events. A source or release row without a download figure means exactly that—there is a verifiable artifact, but no defensible public usage count.

The v0.5.0 tree also carried provider packages that did not leave a current, independently measurable PyPI record. They still belong in the implementation record, but not in a download total.

Tooling around the format

A portable format needs more than parsers. It needs hermetic builds, syntax support, language servers, formatters, and release paths that make each implementation ordinary to use in its own ecosystem.

The two largest Bazel efforts are rules_dart and rules_flutter: toolchain resolution, persistent workers, remote-build execution, IDE aspects, Gazelle integration, application packaging, and version-conflict detection. Smaller adapters make the shared Vitest, Java, and Rust suites participate in the same hermetic build graph. Promptly then carries the format in the other direction, from the build into the editor: lint, format, test, publish, and language-server operations over .prompt files.

Public-source use of Genkit Python

GitHub's code index returned dozens of independent public repositories containing source-level Genkit Python imports after excluding the Genkit organization, the author's own repositories, vendored virtual environments, and repository forks. Search indexes are incomplete and public code is not proof of production deployment, so the table is a representative, manually checked subset rather than a claim about total users.

What the numbers do and do not say

Dotprompt's TypeScript package is the mature distribution channel; its npm traffic is two orders of magnitude larger than the younger Python core. Genkit Python, Dotprompt's new runtime packages, and the editor/Bazel work are earlier on the adoption curve. The useful evidence today is therefore mixed: registry downloads for established packages, release and source records for new ones, and pinned imports for public applications.

The stars and forks on the shared Genkit and Dotprompt repositories describe the whole multi-language projects. They should not be relabeled as Python-only adoption. The same restraint applies to Google's statement that Genkit is used in production: the public source does not establish that those production workloads use the Python SDK.

—-

For the mechanism beneath the adoption record, Everything Is an Action dissects Genkit's architecture from source. If a public Genkit Python or Dotprompt use is missing here, I'd be glad to hear about it.