Exaloop AI
Exaloop AI equips coding agents to build native data applications with Exaloop. It combines API recipes, Codon language guidance, runnable examples and validation tools, covering the path from implementation through integration, measurement and deployment.
Use it to work with Arrow and Parquet, generate Avro and Protobuf records, connect to databases, build native database functions, process Kafka streams, or expose native calculations to Python. The original Codon guidance remains available for Python migration, native NumPy, classes, parallelism and low-level tuning.
The distribution includes the complete bundle under share/exaloop-ai. Its
assistant-compatible entry point uses the Agent Skills format; that is how the
resources are discovered, not the limit of their coverage. Activating the bundle
does not install SDKs or grant an agent access to databases, brokers or credentials.
Build With Your Agent
| Task | Guidance supplied |
|---|---|
| Data pipelines | Exact Arrow types, nullable values, ownership, file formats and native serialization |
| Database integration | Typed clients, parameterized SQL, transactions, streaming results and native host entry points |
| Kafka applications | Delivery reports, backpressure, offsets, transactions and Avro schema-registry codecs |
| Python integration | Interpreter selection, ABI-named modules, installable wheels and bundled runtimes |
| Performance work | Native profiling, representative PGO training, holdout validation and measured comparisons |
| Delivery | Project targets, SDK selection, runtime packaging, host requirements and clean-destination checks |
The entry point routes the assistant to the relevant reference. Detailed APIs are loaded on demand, keeping unrelated SDK material out of the current task.
Install and Activate
For the default installation, copy the supplied bundle into a supported project skill directory:
mkdir -p .github/skills
cp -R "${EXALOOP_HOME:-$HOME/.exaloop}/current/share/exaloop-ai" .github/skills/exaloop
This example uses GitHub Copilot's project skill directory. Other clients may use
.agents/skills/exaloop. The destination should be empty when installing; an
existing copy can be replaced deliberately when updating. The complete directory
is needed because SKILL.md refers to the accompanying examples and references.
Activate it with /exaloop where your client supports that command, or ask the
assistant to use Exaloop AI. Check your client's loaded-resource indicators to
confirm that the entry is active. Keep the complete directory together.
If you previously installed the Codon-only bundle, install this version under
exaloop and retire the old activation deliberately. Older distributions may
still contain share/codon-skill; they do not contain the expanded API references.
The existing documentation URL remains valid.
Example requests:
Use Exaloop AI to build a Parquet pipeline. Preserve nullable columns, choose explicit native types, and package the executable with its runtime libraries.
Write a PostgreSQL client using bound parameters and explicit transaction commits. Keep credentials in private configuration and use a development database.
Build an Avro Kafka worker. Preserve keys and tombstones, and choose an offset and delivery policy that remains correct when the worker restarts.
Package this calculation as a Python wheel for my virtual environment. Verify its exported calls after installation without the development toolchain.
Use Pythonic mode to port this numeric stage. Preserve array mutation and validate it under CPython and Codon.
Use Low-Level mode for this fixed-size parser. Prove bounds, lifetime, and overlap assumptions, then compare it with release Pythonic Codon.
Choose a Mode
| Mode | Approach | Best suited to |
|---|---|---|
| Pythonic, default | Familiar typed Python structures and native NumPy for dense numerical work | Porting an application while keeping it readable and preserving its behavior |
| Low-Level | Explicit data layouts, bounded buffers, pointers, SIMD, and selected C/LLVM operations | A measured bottleneck where more control over memory or computation is worthwhile |
Pythonic mode is a useful starting point. Low-Level mode asks for more explicit reasoning about bounds, lifetime, alignment, and concurrency, as well as a comparison against the simpler implementation.
Tell the assistant how you plan to run the result: as a native executable, as a Python-hosted JIT or extension, or with explicit Python interoperability. That choice affects how much code can remain native and where conversion costs occur. Codon's native NumPy is bundled with the compiler; CPython NumPy is a separate reference implementation used by some tests.
Resource Map
| Resource in the installed bundle | Contents |
|---|---|
SKILL.md | The guidance the assistant reads first, including how to choose an implementation approach |
references/exaloop-workflows.md | CLI targets, SDK selection, Python wheels, profiling, PGO and deployment |
references/exaloop-data.md | Arrow ownership and nulls, Avro/Protobuf generation and wire contracts |
references/exaloop-databases.md | SQL/Redis client APIs, transactions and native host entry points |
references/exaloop-streaming.md | Kafka delivery, offset commits, transactions and registry framing |
references/compatibility.md, migration.md | Semantic differences and migration boundaries |
references/classes.md | Reference versus tuple/value records and class layouts |
references/low-level.md, parallel-interop.md | Unsafe memory, SIMD, C ABI, threads, and GPU boundaries |
references/stdlib.md, stdlib-catalog.json | Native module guidance and revision-pinned inventory |
references/validation.md, sources.md | Testing and measurement methods, with upstream references |
examples/ | Nine probes covering Pythonic/NumPy/low-level, classes, semantics, stdlib, and parallelism |
assets/delivery-template.md | A template for summarizing changes, tests, measurements, and limitations |
The catalog records the compiler revision it describes. The included compatibility reference and the active compiler's documentation provide the relevant API details.
Command-Line Tools
The local tools require Python 3.9 or newer. They use the standard library and do not install packages, upload data, or rewrite the application. With the bundle installed in the location above:
python3 .github/skills/exaloop/scripts/audit.py path/to/application.py
mkdir -p build
python3 .github/skills/exaloop/scripts/validate.py --codon /path/to/codon \
--report build/ai-validation.json
| Tool | Arguments and outputs |
|---|---|
audit.py | One or more Python file/directory paths; JSON heuristic review findings, no target execution/import |
validate.py | Required --codon; --timeout per command (180s); optional --report PATH; --benchmark, --size (200000), --repeats (5) |
The core validator builds nine Codon examples in temporary directories and runs them in
debug and release mode with -numerics=py. It checks outputs and edge cases,
emits LLVM IR, and writes a final report when --report is supplied. Create the
report's parent directory before running it. Optional CPython NumPy comparisons
are marked as skipped when unavailable; native NumPy tests still run.
Those checks do not certify SDK integrations or external services. The product references call for actual API calls, database-host loading, broker delivery, wheel installation or clean-destination execution as appropriate. Compile-only checks and unrun services must be reported explicitly. The retained evaluation catalog covers Codon tasks; it is not a product-wide claim about agent success.
Use --benchmark to include process-level timings. Those measurements include
startup, initialization, and output, but exclude compilation, so they are a
starting point rather than an isolated kernel benchmark.
Documentation for Assistants
Each published guide is also available as Markdown through its page actions. The documentation index lists the available guides. An assistant's output still needs the same code review, tests, and production validation as other application changes.