The OpenSAS community, although currently focused on the mandatory subset of SAS used in clinical trials, is always keeping an eye on expanding its feature support to achieve full compatibility with other industry use cases.

It reads and writes .sas7bdat and .xpt files and runs the DATA step, the macro language, PROC SQL and the reporting and data-management PROCs. Built on production clinical-trial (CDISC) pipelines.

• Audit compliant. Zero-dependency static binary that runs ubiquitously across Linux, macOS, Windows, WebAssembly.
• Checked against SAS® 9.4. OpenSAS output datasets have been diffed against official runtime output. For what it supports, the bytes match.
• Fails loud. Unsupported language feature stops with a dedicated error code (2) and line number so you or your agent can differentiate from programming error (1). Zero false positives.

Quick start

Grab a binary from the releases page (Linux arm64 and x86_64, Windows x86_64, macOS arm64), or build it with Zig 0.16.0:

git clone https://github.com/kirha-ai/opensas && cd OpenSAS
zig build -Doptimize=ReleaseFast        # binary at zig-out/bin/sas

Write a program:

data vitals;
  input usubjid $ visit $ sysbp diabp;
  map = round(diabp + (sysbp - diabp) / 3, 0.1);
  datalines;
S001 BASELINE 128 82
S001 WEEK4    121 79
S002 BASELINE 142 91
S002 WEEK4    135 88
;
run;

proc sort data=vitals; by usubjid descending visit; run;

proc print data=vitals noobs;
  where map > 100;
run;

Run it:

$ sas vitals.sas
usubjid   visit      sysbp   diabp     map

S002      WEEK4        135      88   103.7
S002      BASELINE     142      91     108

sas --sasautos ./macros program.sas adds an autocall macro library. A libname points to a directory; members resolve as .sas7bdat, .xpt or .csv, and datasets the program writes land there in the same formats. tests/programs/dm is a complete example: a raw demographics dataset mapped to the SDTM DM domain, with the expected output.

Exit codes

A script or an agent can route on the code without parsing the log: 1 is your program, 2 is our gap.

In the browser

The interpreter compiles to wasm32-wasi and runs in the page. Programs and data never leave your machine. The hosted version is at kirha.com/playground/opensas; to serve it yourself:

make wasm                        # builds web/sas.wasm
python3 -m http.server -d web    # open http://localhost:8000

How it is verified

Three suites run in CI on every push and before every release.

• Unit tests. 909 test blocks next to the code they cover.
• Conformance corpus. 1,894 one-feature programs in tests/corpus/. Each program runs and its listing is diffed against the expected output.
• End-to-end programs. Whole programs in tests/programs/, datasets in and datasets out. Every output column and value is diffed against the golden dataset.

The .sas7bdat and .xpt readers and writers are round-tripped against files written by SAS® 9.4. Fixtures use invented data or published samples such as the CDISC pilot study.

What's supported

DATA step. Assignment, conditionals and select, all forms of do loops, where, arrays, retain, lag and dif, multiple output targets, drop/keep/rename, length/label/attrib, hash objects, and formatted, column and list input.

BY-group processing. first. and last. flags, merge with in=, set concatenation, and the dataset options (keep=, drop=, rename=, where=, firstobs=, obs=).

Macro language. Nested macros with keyword and positional parameters, %let/%global/%local, %if and %do, %eval and %sysevalf, %sysfunc, the quoting functions, && indirection, %include, autocall libraries, and CALL SYMPUT/SYMPUTX resolved at step boundaries.

Functions and formats. 468 built-in functions with missing-value semantics (full list). Numeric, character and date/time formats and informats, PERCENT, and user formats through PROC FORMAT.

Procedures. PRINT, SORT, MEANS and SUMMARY, FREQ, UNIVARIATE, TRANSPOSE, REPORT, TABULATE, SQL, DATASETS, COMPARE, CONTENTS, FORMAT, IMPORT and EXPORT (CSV, delimited, XLSX), DELETE.

Dataset I/O. Read and write .sas7bdat and XPORT .xpt, plus CSV-backed libraries.

The grammar is catalogued production by production in docs/sas9.4.ebnf, with an (* opensas *) mark on each implemented one.

Not there yet. Modeling and survival procedures (GLM, MIXED, LOGISTIC, LIFETEST, ...) and graphics procedures (GPLOT, SGPLOT, ...). They exit with code 2. ODS statements for destinations other than the listing are parsed and ignored.

Development

zig build                    # debug binary
zig build test               # unit tests
zig build corpus             # conformance corpus
zig build programs           # end-to-end programs
make arm amd windows macos   # cross-compile release binaries into dist/

src/             lexer, parser, DATA step, macro processor, PROCs, dataset I/O
tests/corpus/    one-feature .sas programs with expected listings
tests/programs/  end-to-end programs with input and golden datasets
docs/            grammar and function trackers, design decisions
web/             browser demo

Contributing

The most useful contribution is a program that runs in SAS® 9.4 but not here. Open an issue with the smallest .sas that reproduces it, the expected output, and what OpenSAS produces along with its exit code.

Pull requests are welcome. Every non-trivial change ships with a test: a test block next to the code, or a fixture in tests/corpus/ with its expected output. House rules live in CLAUDE.md, settled design decisions in docs/decisions.md.

About

OpenSAS is a research project by KIRHA. Coding agents wrote the implementation from the public SAS® 9.4 documentation. Correctness is judged against reference output through the suites above, not against the agents' own assessment. How the team was organised, what worked and what did not is written up at kirha.com/research/opensas.

License

Apache 2.0.

SAS® is a registered trademark of SAS Institute Inc. This project is not affiliated with, endorsed by, or sponsored by SAS Institute Inc.