pyffi

Go bindings for CPython via purego — no Cgo required.

rt, _ := pyffi.New()
defer rt.Close()

rt.Exec(`x = 1 + 2`)
result, _ := rt.Eval("x * 10")
defer result.Close()
fmt.Println(result.Int64()) // 30

Features

No Cgo — pure Go, uses purego for FFI. CGO_ENABLED=0 builds work
Auto-GIL — all methods are goroutine-safe out of the box
Full type conversion — bool, int, float, string, bytes, list, tuple, dict, set
Async support — RunAsync, CallAsync, EventLoop for asyncio integration
Callbacks — register Go functions callable from Python (with kwargs)
Free-threaded Python — automatic detection of Python 3.13t+ builds
uv integration — auto-detect, auto-install dependencies, project venv support
Code generation — pyffi-gen generates type-safe Go bindings from Python modules
Platform support — macOS, Linux, and Windows

Install

go get github.com/i2y/pyffi

Python 3.12+ is needed at runtime. pyffi auto-detects system Python, Homebrew, and uv-managed installations. If Python is not installed:

# Recommended: install uv, then let pyffi auto-detect
curl -LsSf https://astral.sh/uv/install.sh | sh
uv python install 3.14

Quick Start

Execute and Evaluate

rt, _ := pyffi.New()
defer rt.Close()

// Execute statements
rt.Exec(`
def greet(name):
    return f"Hello, {name}!"
`)

// Evaluate expressions
result, _ := rt.Eval(`greet("World")`)
defer result.Close()
s, _ := result.GoString() // "Hello, World!"

Resource Management (Close)

pyffi wraps Python objects as *Object values that hold a reference to the underlying CPython object. These must be released with Close() to avoid memory leaks:

Type	Close required?	Why
`*Runtime`	Yes (`defer rt.Close()`)	Calls `Py_Finalize` to shut down the interpreter
`*Object`	Yes (`defer obj.Close()`)	Calls `Py_DecRef` to release the Python reference

Every method that returns *Object — Eval, Import, Attr, Call, GetItem, Iter().Next(), etc. — requires the caller to close the result. A GC finalizer exists as a safety net, but explicit Close() is strongly recommended.

Primitive extraction methods (Int64, Float64, GoString, Bool, GoSlice, GoMap, GoValue) return plain Go values that do not need closing.

Import Modules

math, _ := rt.Import("math")
defer math.Close()

pi := math.Attr("pi")
defer pi.Close()
f, _ := pi.Float64() // 3.141592653589793

Collections

// Create
list, _ := rt.NewList(1, "two", 3.0)
dict, _ := rt.NewDict("name", "Go", "year", 2009)
tuple, _ := rt.NewTuple("a", "b", "c")
set, _ := rt.NewSet(1, 2, 3)

// Access
item, _ := list.GetItem(0)      // list[0]
list.SetItem(0, 99)             // list[0] = 99
list.DelItem(0)                 // del list[0]
ok, _ := list.Contains(2)       // 2 in list
n, _ := list.Len()              // len(list)
r, _ := list.Repr()             // repr(list)

// Iterate
iter, _ := list.Iter()
defer iter.Close()
for {
    item, _ := iter.Next()
    if item == nil { break }
    defer item.Close()
    // ...
}

// Compare
eq, _ := a.Equals(b)            // a == b
lt, _ := a.Compare(b, pyffi.PyLT)  // a < b

Type Conversions

Go → Python	Python → Go
`bool` → `bool`	`bool` → `bool`
`int`, `int8`–`int64` → `int`	`int` → `int64`
`uint`, `uint8`–`uint64` → `int`	`float` → `float64`
`float32`, `float64` → `float`	`str` → `string`
`string` → `str`	`bytes`, `bytearray` → `[]byte`
`[]byte` → `bytes`	`list`, `tuple`, `set` → `[]any`
`[]any` → `list`	`dict` → `map[string]any`
`map[string]any` → `dict`	`None` → `nil`
`nil` → `None`

Other Python types (functions, classes, instances, modules, etc.) are returned as *Object and can be manipulated via Attr(), Call(), GetItem(), etc:

// Classes: Call() instantiates
cls := mod.Attr("MyClass")
defer cls.Close()
instance, _ := cls.Call("arg1", 42)  // MyClass("arg1", 42)
defer instance.Close()

// Instance methods and attributes
name, _ := instance.Attr("name").GoString()
result, _ := instance.Attr("method").Call()
defer result.Close()

// Functions: first-class objects
fn := mod.Attr("some_function")
defer fn.Close()
result, _ := fn.Call(args...)

Goroutine Safety

All methods automatically acquire the GIL. No manual management needed:

var wg sync.WaitGroup
for i := range 10 {
    wg.Add(1)
    go func(n int) {
        defer wg.Done()
        obj := rt.FromInt64(int64(n))
        defer obj.Close()
        // Safe from any goroutine
    }(i)
}
wg.Wait()

For batching (reduces GIL overhead):

rt.WithGIL(func() error {
    rt.Exec("a = 1")
    rt.Exec("b = 2")
    rt.Exec("c = a + b")
    return nil
})

Async Python

// Synchronous
result, _ := rt.RunAsync("fetch_data('https://example.com')")

// Non-blocking (background goroutine)
ch := rt.RunAsyncGo("fetch_data('https://example.com')")
ar := <-ch // pyffi.AsyncResult{Value, Err}

// Call async functions directly
fn := mod.Attr("async_func")
result, _ := fn.CallAsync(arg1, arg2)

Callbacks

rt.RegisterFunc("add", func(a, b int) int {
    return a + b
})

// With keyword arguments
rt.RegisterFunc("greet", func(name string, kw map[string]any) string {
    greeting := "hello"
    if g, ok := kw["greeting"]; ok {
        greeting = g.(string)
    }
    return greeting + " " + name
})

rt.Exec(`
import go_bridge
print(go_bridge.add(1, 2))               # 3
print(go_bridge.greet("Go", greeting="hi"))  # hi Go
`)

Context Managers

rt.With(resource, func(value *pyffi.Object) error {
    // __enter__ called, value is the result
    // __exit__ called automatically (even on error)
    return nil
})

Python & Dependency Management with uv

pyffi integrates with uv for Python discovery and dependency management. uv is a fast Python package manager written in Rust.

Auto-Detect uv-Managed Python

// Prefer uv-managed Python installations
rt, _ := pyffi.New(pyffi.WithUV())

Inline Dependencies

Automatically creates a hash-based cached venv in ~/.cache/pyffi/venvs/ and installs packages:

rt, _ := pyffi.New(pyffi.Dependencies("numpy", "pandas", "requests"))
defer rt.Close()

rt.Exec(`
import numpy as np
import pandas as pd
print(np.array([1, 2, 3]))
`)

The venv is cached by a hash of the dependency list — subsequent runs skip installation.

Project venv (pyproject.toml)

For projects with a pyproject.toml, use WithUVProject to run uv sync and use the project's venv:

rt, _ := pyffi.New(pyffi.WithUVProject("/path/to/python-project"))
defer rt.Close()

// All project dependencies are available
rt.Exec(`from mypackage import something`)

Explicit Library Path

Skip auto-detection entirely by pointing directly to the Python shared library:

rt, _ := pyffi.New(pyffi.WithLibraryPath("/usr/lib/libpython3.14.so"))

This is useful in containers where the Python location is known at build time.

Code Generation

pyffi-gen generates type-safe Go bindings from Python modules by introspecting them at build time:

# Install
go install github.com/i2y/pyffi/cmd/pyffi-gen@latest

# Generate bindings for a Python module
pyffi-gen --module numpy --out ./gen/numpypkg --dependencies numpy

# Preview without writing files
pyffi-gen --module json --dry-run

# Use a config file
pyffi-gen --config pyffi-gen.yaml

# Initialize a new project with config scaffolding
pyffi-gen init --module numpy,pandas

The generated code provides a low-level typed Module wrapper. For production use, create a higher-level Go package that wraps the generated code with idiomatic APIs:

yourpkg/
├── internal/sdk/       # Generated by pyffi-gen (DO NOT EDIT)
│   └── sdk.go
├── yourpkg.go          # Your idiomatic Go API wrapping internal/sdk
├── options.go          # Option types
└── ...

You don't need to wrap every generated function — pick and choose what to expose. You can also bypass the generated bindings entirely and use rt.Exec() / rt.Eval() with Python code strings for cases where that's simpler (e.g., variadic expression arguments). Mixing both approaches in the same package is fine. The polarsgo package in this repository is a practical example — it uses the generated internal/sdk/ for methods with fixed signatures (Head, Tail, Join, Sort, etc.) while using inline Python code for expression-based methods (Filter, WithColumns, GroupBy).

Docker Deployment

pyffi works well in Docker with a multi-stage build. Since pyffi uses purego (no Cgo), the Go binary can be built with CGO_ENABLED=0:

# Stage 1: Build Go binary (no Python needed)
FROM golang:1.26 AS builder
WORKDIR /app
COPY go.mod go.sum ./
RUN go mod download
COPY . .
RUN CGO_ENABLED=0 go build -o /app/myapp ./cmd/myapp

# Stage 2: Runtime with Python
FROM python:3.14-slim
RUN pip install uv
COPY --from=builder /app/myapp /usr/local/bin/myapp

# Option A: Use pyproject.toml
COPY pyproject.toml uv.lock ./
RUN uv sync
# In Go: pyffi.New(pyffi.WithUVProject("."))

# Option B: Or let pyffi.Dependencies() install at first run
# In Go: pyffi.New(pyffi.Dependencies("numpy", "pandas"))

CMD ["myapp"]

Optimized: No uv at Runtime

For smaller images, install dependencies at build time and use WithLibraryPath to skip uv at runtime:

FROM python:3.14-slim AS python-deps
RUN pip install uv
COPY pyproject.toml uv.lock ./
RUN uv sync

FROM python:3.14-slim
COPY --from=python-deps /.venv /.venv
COPY --from=builder /app/myapp /usr/local/bin/myapp
# No uv needed at runtime
# In Go: pyffi.New(pyffi.WithLibraryPath("/usr/local/lib/libpython3.14.so"))
# Then add /.venv/lib/python3.14/site-packages to sys.path via Exec
CMD ["myapp"]

Wrapper Packages

pyffi-powered Go bindings for popular Python libraries. Each is an independent Go module — install only what you need.

polarsgo — DataFrames

Fast DataFrame operations from Go using Polars. Filter, sort, join, group, aggregate, LazyFrame optimization, and SQL queries.

go get github.com/i2y/pyffi/polarsgo

pl, _ := polarsgo.New()
defer pl.Close()

df, _ := pl.ReadCSV("data.csv")
result, _ := df.Filter("col('age') > 30").Sort("age", true)
fmt.Println(result)

// SQL queries
sqlResult, _ := pl.SQL("SELECT dept, AVG(age) FROM t GROUP BY dept", map[string]*polarsgo.DataFrame{"t": df})

See the polarsgo README for the full API including LazyFrame, Join, GroupBy, and more.

sbert — Sentence Embeddings

Generate semantic embeddings using 15,000+ sentence-transformers models.

go get github.com/i2y/pyffi/sbert

model, _ := sbert.New("all-MiniLM-L6-v2")
defer model.Close()

embeddings, _ := model.Encode([]string{"Hello world", "Go is great"})
sim, _ := model.Similarity(embeddings, embeddings)

See the sbert README.

hfpipe — Hugging Face Pipelines

Local ML inference — text generation, classification, summarization, and more.

go get github.com/i2y/pyffi/hfpipe

pipe, _ := hfpipe.New("text-classification", "distilbert/distilbert-base-uncased-finetuned-sst-2-english")
defer pipe.Close()

results, _ := pipe.Run("I love this movie!")  // [{label:POSITIVE score:0.9999}]

See the hfpipe README.

dspygo — DSPy

Program (not prompt) language models with typed signatures, pipelines, and automatic prompt optimization.

go get github.com/i2y/pyffi/dspygo

client, _ := dspygo.New(dspygo.WithLM("openai/gpt-4o-mini"))
defer client.Close()

classify := client.PredictSig(dspygo.Signature{
    Doc: "Classify sentiment.",
    Inputs:  []dspygo.Field{{Name: "sentence", Type: "str"}},
    Outputs: []dspygo.Field{{Name: "sentiment", Type: `Literal["positive", "negative", "neutral"]`}},
})
result, _ := classify.Call(dspygo.KV{"sentence": "I love it!"})

See the dspygo README.

outlines — Structured Generation

Constrained decoding — guarantee LLM output matches a JSON schema, regex, or choice set.

go get github.com/i2y/pyffi/outlines

model, _ := outlines.NewOllama("llama3.2")
defer model.Close()

result, _ := model.PydanticJSON("Generate a user profile.", "Profile", map[string]string{"name": "str", "age": "int"})

See the outlines README.

diffusersgo — Image Generation

Generate images from text using Hugging Face Diffusers (Stable Diffusion, FLUX, etc.).

go get github.com/i2y/pyffi/diffusersgo

pipe, _ := diffusersgo.New("stable-diffusion-v1-5/stable-diffusion-v1-5",
    diffusersgo.WithDevice("mps"),
    diffusersgo.WithDtype("float16"),
)
defer pipe.Close()

img, _ := pipe.TextToImage("A cat in space, oil painting")
img.Save("cat.png")

See the diffusersgo README.

smolagentsgo — Lightweight Agents

Build agents with smolagents that write Python code to orchestrate tools and solve multi-step tasks.

go get github.com/i2y/pyffi/smolagentsgo

client, _ := smolagentsgo.New(
    smolagentsgo.WithLiteLLM("anthropic/claude-3-haiku-20240307", apiKey),
)
defer client.Close()

result, _ := client.Run("What is 15 * 23?")
fmt.Println(result) // 345

See the smolagentsgo README.

casdk — Claude Agent SDK

Go wrapper for the Claude Agent SDK with hooks, plugins, and in-process MCP tools.

go get github.com/i2y/pyffi/casdk

client, _ := casdk.New()
defer client.Close()

for msg, err := range client.Query(ctx, "What is 2+2?", casdk.WithMaxTurns(1)) {
    if err != nil { log.Fatal(err) }
    fmt.Println(msg.Text())
}

See the casdk README.

datasetsgo — Hugging Face Datasets

Access 200,000+ datasets on the Hugging Face Hub from Go.

go get github.com/i2y/pyffi/datasetsgo

client, _ := datasetsgo.New()
defer client.Close()

ds, _ := client.Load("rotten_tomatoes")
train := ds.Split("train")
fmt.Println(train.Len())       // 8530
row, _ := train.Row(0)
fmt.Println(row["text"])

See the datasetsgo README.

peftgo — LoRA Fine-Tuning

Fine-tune large models with PEFT (LoRA, etc.) from Go.

go get github.com/i2y/pyffi/peftgo

model, _ := peftgo.NewModel("meta-llama/Llama-3-8B", peftgo.WithDevice("auto"))
defer model.Close()

model.ApplyLoRA(peftgo.LoRAConfig{Rank: 16, Alpha: 32})
trainable, total, pct := model.TrainableParams()
fmt.Printf("Trainable: %d / %d (%.2f%%)\n", trainable, total, pct)
model.SaveAdapter("./my-adapter")

See the peftgo README.

Known Limitations

Process exit SIGSEGV: When using Python libraries with background threads (PyTorch, datasets, etc.), a SIGSEGV may occur during process shutdown. This does not affect program correctness — it only happens after main() returns and the OS is about to reclaim all resources anyway. Long-running servers are unaffected since the issue only occurs at process exit.
Py_Finalize instability: CPython's Py_Finalize does not work reliably with all extension modules. Use pyffi.WithSkipFinalize() if you encounter issues. Resources are reclaimed by the OS on process exit.

For LLM Agents

This project includes AgentSkills for AI coding assistants:

skills/pyffi/SKILL.md — pyffi usage guide
casdk/skills/casdk/SKILL.md — casdk usage guide
llms.txt — LLM-friendly API index
llms-full.txt — Full documentation for LLM context

Build & Test

make ci          # fmt + vet + build + test
make test        # verbose tests
make bench       # benchmarks
make lint        # staticcheck (if installed)

License

MIT

Name		Name	Last commit message	Last commit date
Latest commit History 28 Commits
.github/workflows		.github/workflows
casdk		casdk
cmd/pyffi-gen		cmd/pyffi-gen
datasetsgo		datasetsgo
diffusersgo		diffusersgo
dspygo		dspygo
examples		examples
hfpipe		hfpipe
internal/gen		internal/gen
oasdk		oasdk
outlines		outlines
peftgo		peftgo
polarsgo		polarsgo
sbert		sbert
skills/pyffi		skills/pyffi
smolagentsgo		smolagentsgo
.gitignore		.gitignore
LICENSE		LICENSE
Makefile		Makefile
README.md		README.md
access_test.go		access_test.go
async_test.go		async_test.go
benchmark_test.go		benchmark_test.go
bind.go		bind.go
bind_test.go		bind_test.go
bytes_test.go		bytes_test.go
callback.go		callback.go
callback_test.go		callback_test.go
compare_test.go		compare_test.go
construct_test.go		construct_test.go
context_test.go		context_test.go
convert.go		convert.go
detect.go		detect.go
detect_darwin.go		detect_darwin.go
detect_linux.go		detect_linux.go
detect_windows.go		detect_windows.go
dlopen_unix.go		dlopen_unix.go
dlopen_windows.go		dlopen_windows.go
errors.go		errors.go
errors_test.go		errors_test.go
eval_test.go		eval_test.go
eventloop.go		eventloop.go
eventloop_test.go		eventloop_test.go
example_test.go		example_test.go
gil.go		gil.go
go.mod		go.mod
go.sum		go.sum
kw.go		kw.go
llms-full.txt		llms-full.txt
llms.txt		llms.txt
object.go		object.go
object_test.go		object_test.go
python.go		python.go
python_test.go		python_test.go
set_test.go		set_test.go
shared_test.go		shared_test.go
tuple_test.go		tuple_test.go
uint64_test.go		uint64_test.go
uv.go		uv.go
uv_cache.go		uv_cache.go
uv_test.go		uv_test.go
version_test.go		version_test.go

Folders and files

Latest commit

History

Repository files navigation

pyffi

Features

Install

Quick Start

Execute and Evaluate

Resource Management (Close)

Import Modules

Collections

Type Conversions

Goroutine Safety

Async Python

Callbacks

Context Managers

Python & Dependency Management with uv

Auto-Detect uv-Managed Python

Inline Dependencies

Project venv (pyproject.toml)

Explicit Library Path

Code Generation

Docker Deployment

Optimized: No uv at Runtime

Wrapper Packages

polarsgo — DataFrames

sbert — Sentence Embeddings

hfpipe — Hugging Face Pipelines

dspygo — DSPy

outlines — Structured Generation

diffusersgo — Image Generation

smolagentsgo — Lightweight Agents

casdk — Claude Agent SDK

datasetsgo — Hugging Face Datasets

peftgo — LoRA Fine-Tuning

Known Limitations

For LLM Agents

Build & Test

License

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages