Skip to content

A2UI Express developer guide

A2UI Express is a compact, token-efficient declarative syntax designed for dynamic generative user interfaces. It acts as an intermediate, highly compressed notation that local on-device or remote Gemini models generate. A host-side compiler parses this syntax and compiles it into standard A2UI v1.0 wire protocol payloads.


Environment Variable Gate

To prevent pollution of the stable v1.0 baseline workspace during development, all A2UI Express code imports and command-line utilities are gated behind the environment variable A2UI_EXPRESS_ENABLED.

You must prepend A2UI_EXPRESS_ENABLED=true to all python command invocations.


Compiling dynamic user interfaces

By default, the testing and evaluation tools are designed to run using remote Gemini models via the standard Google GenAI SDK.

Running inference and validation (Gemini API)

To run inference and compiler validation using a standard remote Gemini model (e.g., gemini-3.1-flash-lite):

  1. Set your API Key:
export GEMINI_API_KEY="your_gemini_api_key_here"
  1. Navigate and Execute:
# Navigate to the express directory
cd specification/proposals/express

# Run inference against Gemini API (using the a2ui_agent project environment)
A2UI_EXPRESS_ENABLED=true uv run --project ../../../python/a2ui_agent scripts/run_inference.py \
  ../../../catalogs/basic/v1/examples/01_flight-status.json \
  --model gemini-3.1-flash-lite

The script will:

  1. Extract the target component structure from the JSON example.
  2. Compile the active catalog schema definitions into plain-text positional signatures.
  3. Query the Gemini API using the prompt contract system instructions.
  4. Compile the returned A2UI Express DSL back into pretty-printed, standard A2UI v1.0 JSON.
  5. Validate the final component tree structure, checking parent-child references and data pointer paths.

CLI utility reference

The express package provides standalone developer scripts in specification/proposals/express/scripts/. Each script dynamically adjusts python paths during execution, allowing them to run directly from any directory when invoked with the appropriate project environment.

Direct prompt generation

Generate the model prompt contract, containing positional component signatures and rules compiled from the active catalog schema:

A2UI_EXPRESS_ENABLED=true uv run --project ../../../python/a2ui_agent scripts/run_prompt_generator.py --catalog ../../../catalogs/basic/v1/catalog.json

Plain DSL compiler

Compile an offline A2UI Express DSL file directly into standard pretty-printed v1.0 JSON:

A2UI_EXPRESS_ENABLED=true uv run --project ../../../python/a2ui_agent scripts/run_compiler.py \
  path/to/sample.a2ui \
  --surface-id "dashboard_surface"

--surface-id is the surface used when the DSL has no surface(...) statement.

JSON-to-Express decompiler

Convert a standard A2UI v1.0 JSON example back into compact A2UI Express code. The script takes the components of the example's first updateComponents message, wraps them in a createSurface for the basic catalog, and prints the resulting surface(...) block:

A2UI_EXPRESS_ENABLED=true uv run --project ../../../python/a2ui_agent scripts/run_decompiler.py ../../../catalogs/basic/v1/examples/01_flight-status.json

Regenerate documentation examples

Regenerate the dynamic express_dsl_examples.md file containing the system prompt contract and compiled weather forecast examples directly from the active code:

A2UI_EXPRESS_ENABLED=true uv run --project ../../../python/a2ui_agent scripts/recreate_dsl_examples.py