swaplm 0.3.0

Universal Python SDK for LLM providers. One interface, every model.

▄▄▄▄▄▄▄ ▄▄▄▄ ▄▄▄ ▄▄▄▄ ▄▄▄▄ ▄▄▄▄▄▄▄ ▄▄▄ ▄▄▄ ▄▄▄ █████▀▀▀ ▀███ ███ ███▀ ▄██▀▀██▄ ███▀▀███▄ ███ ████▄ ▄████ ▀████▄ ███ ███ ███ ███ ███ ███▄▄███▀ ███ ███▀████▀███ ▀████ ███▄▄███▄▄███ ███▀▀███ ███▀▀▀▀ ███ ███ ▀▀ ███ ███████▀ ▀████▀████▀ ███ ███ ███ ████████ ███ ███

The universal Python SDK for every LLM provider.

One interface. 16 providers. Zero vendor lock-in.

---

Why SwapLM?

Most LLM applications are tightly coupled to a single provider. Switching from Groq to OpenAI to Anthropic means rewriting your API calls, error handling, streaming logic, and auth management. SwapLM eliminates that lock-in.

Write your code once. Swap providers by changing a single string.

from swaplm import chat

# Groq
response = chat(
    provider="groq",
    model="llama-3.3-70b-versatile",
    messages=[{"role": "user", "content": "Hello!"}],
)

# OpenAI — same interface, different provider
response = chat(
    provider="openai",
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello!"}],
)

# Anthropic — still the same interface
response = chat(
    provider="anthropic",
    model="claude-sonnet-4-20250514",
    messages=[{"role": "user", "content": "Hello!"}],
)

---

Features

• 🔌 16 Production Providers — Groq, Anthropic, Google Gemini, OpenAI, OpenRouter, GitHub Models, NVIDIA NIM, Cerebras, SambaNova, Mistral, xAI, DeepInfra, Fireworks, Cloudflare, Perplexity, Cohere.
• ⚡ Sync & Async — chat() for synchronous code, achat() for async/await workflows.
• 📡 Streaming — Real-time token-by-token responses with stream=True. Works sync and async.
• 🛠️ Tool / Function Calling — Define tools with JSON schemas. SwapLM handles provider-specific translation automatically.
• 🔄 Retries with Exponential Backoff — Built-in retry on 5xx, 429 rate limits, and timeouts. Configurable per-request or globally.
• 🧩 Middleware Pipeline — Intercept and modify requests/responses with BaseMiddleware. Supports both sync and async.
• 🪝 Lifecycle Hooks — Listen for before_request, after_request, before_retry, on_error events.
• 🔌 Custom Transport — Replace the HTTP layer entirely with BaseTransport for proxies, caching, mocks, or enterprise security.
• 🔍 Provider & Model Discovery — Query available providers and their models programmatically.
• 🧠 Model Capability Awareness — SDK automatically omits unsupported parameters (e.g., seed, tool_choice) based on model metadata.
• 🔒 Secure Debug Logging — configure(debug=True) redacts API keys and authorization headers.
• 📦 Zero Heavy Dependencies — Only pydantic>=2.11 and httpx>=0.27.0.

---

Installation

pip install swaplm

Requires Python 3.10+.

---

Quick Start

1. Set Your API Key

export GROQ_API_KEY=gsk_...    # or any provider's env var

2. Make Your First Call

from swaplm import chat

response = chat(
    provider="groq",
    model="llama-3.3-70b-versatile",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "What is the meaning of life?"},
    ],
    temperature=0.7,
    max_tokens=256,
)

print(response.content)
# => "The meaning of life is a philosophical question..."

3. Async Version

import asyncio
from swaplm import achat


async def main():
    response = await achat(
        provider="groq",
        model="llama-3.3-70b-versatile",
        messages=[{"role": "user", "content": "Hello!"}],
    )
    print(response.content)


asyncio.run(main())

---

API Reference

chat() / achat() Parameters

Both functions accept the same parameters (keyword-only). achat() is the async equivalent.

Parameter	Type	Default	Description
model	str	required	Model identifier. When provider is set, this is the raw model ID. Otherwise "provider/model" format (e.g., "groq/llama-3.3-70b-versatile").
messages	list[dict | Message]	required	Conversation history. Each dict needs "role" and "content".
provider	str | None	None	Explicit provider slug (e.g., "groq"). When set, bypasses model-string parsing.
stream	bool	False	If True, returns a StreamResponse that yields StreamChunk objects.
api_key	str | None	None	Explicit API key. Overrides the provider's environment variable.
base_url	str | None	None	Override the provider's default API base URL.
max_tokens	int | None	None	Maximum tokens to generate in the response.
temperature	float | None	None	Sampling temperature (0.0–2.0). Higher = more random.
top_p	float | None	None	Nucleus sampling threshold (0.0–1.0).
response_format	dict | None	None	Structured output format. Use {"type": "json_object"} for JSON mode.
tools	list[Tool | dict] | None	None	List of tools the model can call.
tool_choice	str | dict | None	None	Controls tool usage: "auto", "none", "required", or a specific tool dict.
stop	str | list[str] | None	None	Stop sequence(s) that halt generation.
seed	int | None	None	Seed for reproducible outputs (provider-dependent).
timeout	float | None	None	Request timeout in seconds. Falls back to configure() default.
retries	int	0	Number of retries on transient failures (5xx, 429, timeouts).
extra_headers	dict[str, str] | None	None	Additional HTTP headers merged into the request.
provider_options	dict[str, Any] | None	None	Provider-specific options passed through verbatim (escape hatch).

Model String Format

Recommended — explicit provider:

chat(provider="groq", model="llama-3.3-70b-versatile")

Still supported — combined model string:

chat(model="groq/llama-3.3-70b-versatile")

Format	Example	Description
Explicit provider	provider="groq", model="llama-3.3-70b-versatile"	Bypasses model-string parsing. Recommended.
Combined string	"groq/llama-3.3-70b-versatile"	Routes directly to the named provider.
Nested model ID	provider="nvidia", model="deepseek-ai/deepseek-v4-flash"	Handles model IDs with slashes cleanly.
Alias	"llama-3.3-70b-versatile"	Searches all providers. Raises AmbiguousModelError if multiple match.

---

Return Types

ChatResponse

Returned by chat() / achat() when stream=False.

from swaplm import chat

response = chat(
    provider="groq",
    model="llama-3.3-70b-versatile",
    messages=[{"role": "user", "content": "Hi"}],
)

# Convenience shortcuts
response.content  # str | None — text of first choice
response.tool_calls  # list[ToolCall] | None — tool calls from first choice
response.finish_reason  # str | None — "stop", "length", "tool_calls", etc.
response.reasoning  # str | None — reasoning/thinking content (if present)

# Full fields
response.id  # str — unique response ID
response.model  # str — model that generated the response
response.provider  # str — provider slug (e.g., "groq")
response.choices  # list[Choice] — all choices
response.usage  # Usage | None — token counts
response.usage.prompt_tokens  # int
response.usage.completion_tokens  # int
response.usage.total_tokens  # int
response.created  # int | None — Unix timestamp

StreamResponse

Returned by chat() / achat() when stream=True.

from swaplm import chat

stream = chat(
    provider="groq",
    model="llama-3.3-70b-versatile",
    messages=[{"role": "user", "content": "Write a poem."}],
    stream=True,
)

# Sync iteration — use convenience properties
for chunk in stream:
    print(chunk.content, end="", flush=True)

# Access full text after iteration
print(stream.text)  # alias for accumulated_content
print(stream.accumulated_content)  # same thing

# Async iteration
async for chunk in stream:
    print(chunk.content, end="", flush=True)

StreamChunk

Each yielded chunk during streaming. Provides convenience properties that proxy the first choice:

Property	Type	Description
content	str	Incremental text content (empty string if none)
reasoning	str	Incremental reasoning/thinking content (empty string if none)
tool_calls	list[ToolCallDelta] | None	Incremental tool call data
finish_reason	str | None	"stop", "length", "tool_calls", etc.

Full access to underlying structure is still available:

Field	Type	Description
id	str | None	Response ID
choices	list[ChunkChoice] | None	Choice deltas
usage	Usage | None	Token usage (if provided by provider)
model	str | None	Model identifier
provider	str | None	Provider slug

ChunkChoice

Field	Type	Description
index	int	Choice index
delta	ChoiceDelta	Incremental content
finish_reason	str | None	"stop", "length", "tool_calls", etc.

ChoiceDelta

Field	Type	Description
role	str | None	Message role (usually "assistant")
content	str | None	Incremental text content
tool_calls	list[ToolCallDelta] | None	Incremental tool call data

---

Supported Providers

Provider	Slug	Protocol	Env Variable	Free Tier	Streaming	Tools	Reasoning
Groq	groq	OpenAI	GROQ_API_KEY	✅	✅	✅	✅
OpenAI	openai	OpenAI	OPENAI_API_KEY	❌	✅	✅	✅
Anthropic	anthropic	Anthropic	ANTHROPIC_API_KEY	❌	✅	✅	✅
Google Gemini	google	Google	GEMINI_API_KEY	✅	✅	✅	✅
OpenRouter	openrouter	OpenAI	OPENROUTER_API_KEY	✅	✅	✅	✅
GitHub Models	github	OpenAI	GITHUB_TOKEN	✅	✅	✅	✅
NVIDIA NIM	nvidia	OpenAI	NVIDIA_API_KEY	✅	✅	✅	✅
Cerebras	cerebras	OpenAI	CEREBRAS_API_KEY	✅	✅	✅	❌
SambaNova	sambanova	OpenAI	SAMBANOVA_API_KEY	✅	✅	✅	✅
Mistral AI	mistral	OpenAI	MISTRAL_API_KEY	✅	✅	✅	❌
xAI	xai	OpenAI	XAI_API_KEY	❌	✅	✅	❌
DeepInfra	deepinfra	OpenAI	DEEPINFRA_API_KEY	❌	✅	✅	✅
Fireworks AI	fireworks	OpenAI	FIREWORKS_API_KEY	❌	✅	✅	✅
Cloudflare	cloudflare	OpenAI	CLOUDFLARE_API_KEY	✅	✅	✅	❌
Perplexity	perplexity	OpenAI	PERPLEXITY_API_KEY	❌	✅	❌	✅
Cohere	cohere	OpenAI	COHERE_API_KEY	✅	✅	✅	❌

OpenAI-compatible providers share the same protocol. Any provider using the /v1/chat/completions endpoint works automatically.

---

Tool / Function Calling

Define tools with JSON Schema parameters. SwapLM translates them to each provider's native format.

from swaplm import Tool, chat

tools = [
    Tool(
        function={
            "name": "get_weather",
            "description": "Get current weather for a location",
            "parameters": {
                "type": "object",
                "properties": {
                    "location": {"type": "string", "description": "City name"},
                    "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]},
                },
                "required": ["location"],
            },
        }
    )
]

response = chat(
    provider="groq",
    model="llama-3.3-70b-versatile",
    messages=[{"role": "user", "content": "What's the weather in Tokyo?"}],
    tools=tools,
    tool_choice="auto",
)

if response.tool_calls:
    for tc in response.tool_calls:
        print(f"Calling: {tc.function.name}")
        print(f"Args: {tc.function.arguments}")

---

Streaming

Sync Streaming

from swaplm import chat

stream = chat(
    provider="groq",
    model="llama-3.3-70b-versatile",
    messages=[{"role": "user", "content": "Tell me a story"}],
    stream=True,
)

for chunk in stream:
    print(chunk.content, end="", flush=True)

# Full text available after iteration
print("\n\nFull:", stream.text)

Async Streaming

import asyncio
from swaplm import achat


async def main():
    stream = await achat(
        provider="groq",
        model="llama-3.3-70b-versatile",
        messages=[{"role": "user", "content": "Tell me a story"}],
        stream=True,
    )

    async for chunk in stream:
        print(chunk.content, end="", flush=True)


asyncio.run(main())

---

Configuration

Set global defaults that apply to all requests:

from swaplm import configure, get_config, reset_config

configure(
    timeout=30.0,  # Request timeout in seconds
    retries=3,  # Retry on 5xx/429/timeout
    retry_delay=0.5,  # Initial retry delay (exponential backoff)
    max_tokens=1024,  # Default max tokens for all requests
    temperature=0.7,  # Default temperature
    top_p=0.9,  # Default top_p
    debug=True,  # Enable debug logging (redacts secrets)
    logging=True,  # Enable info-level logging
)

# Read current config
cfg = get_config()
print(cfg.timeout, cfg.retries)

# Reset to defaults
reset_config()

Config Field	Type	Default	Description
timeout	float | None	None	Global request timeout in seconds.
retries	int	0	Global retry count for transient failures.
retry_delay	float	0.5	Initial delay between retries (doubles each attempt).
max_tokens	int | None	None	Default max tokens if not specified per-request.
temperature	float | None	None	Default temperature if not specified per-request.
top_p	float | None	None	Default top_p if not specified per-request.
debug	bool	False	Enable debug logging with secret redaction.
logging	bool	False	Enable info-level request logging.
transport	BaseTransport | None	None	Custom transport implementation (see below).

---

Custom Transport

Replace the HTTP layer for proxies, caching, mocks, or enterprise security:

from collections.abc import AsyncIterator, Iterator
from typing import Any

from swaplm import BaseTransport, chat, configure, reset_config


class MyTransport(BaseTransport):
    def send(self, method, url, *, headers=None, json=None, timeout=None, retries=0):
        print(f" intercepted: {url}")
        return 200, {"choices": [{"message": {"role": "assistant", "content": "Intercepted!"}}]}

    async def asend(self, method, url, *, headers=None, json=None, timeout=None, retries=0):
        return self.send(method, url, headers=headers, json=json, timeout=timeout, retries=retries)

    def send_stream(self, method, url, *, headers=None, json=None, timeout=None):
        yield {"choices": [{"delta": {"content": "Stream intercepted"}}]}

    async def asend_stream(self, method, url, *, headers=None, json=None, timeout=None):
        yield {"choices": [{"delta": {"content": "Async stream intercepted"}}]}


configure(transport=MyTransport())

response = chat(
    provider="groq",
    model="llama-3.3-70b-versatile",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(response.content)  # => "Intercepted!"

reset_config()

BaseTransport Interface

class BaseTransport(ABC):
    def send(self, method, url, *, headers, json, timeout, retries) -> tuple[int, dict]: ...
    async def asend(self, method, url, *, headers, json, timeout, retries) -> tuple[int, dict]: ...
    def send_stream(self, method, url, *, headers, json, timeout) -> Iterator[dict]: ...
    async def asend_stream(self, method, url, *, headers, json, timeout) -> AsyncIterator[dict]: ...
    def close(self) -> None: ...
    async def aclose(self) -> None: ...

---

Middleware

Intercept and transform requests and responses in a pipeline:

from swaplm import (
    BaseMiddleware,
    ChatRequest,
    ChatResponse,
    add_middleware,
    chat,
    reset_middlewares,
)


class LoggingMiddleware(BaseMiddleware):
    def process_request(self, request: ChatRequest) -> ChatRequest:
        print(f"[REQ] {request.model} | {len(request.messages)} messages")
        return request

    def process_response(self, response: ChatResponse) -> ChatResponse:
        print(f"[RES] {response.provider} | {response.usage.total_tokens} tokens")
        return response


add_middleware(LoggingMiddleware())

response = chat(
    provider="groq",
    model="llama-3.3-70b-versatile",
    messages=[{"role": "user", "content": "Hello!"}],
)
# Prints:
# [REQ] groq/llama-3.3-70b-versatile | 1 messages
# [RES] groq | 42 tokens

reset_middlewares()

BaseMiddleware Methods

Method	Description
process_request(request)	Modify request before execution (sync).
aprocess_request(request)	Modify request before execution (async).
process_response(response)	Modify response after execution (sync).
aprocess_response(response)	Modify response after execution (async).
process_error(error)	Intercept exceptions (sync).
aprocess_error(error)	Intercept exceptions (async).

Pipeline Management

from swaplm import add_middleware, remove_middleware, get_middlewares, reset_middlewares

add_middleware(my_middleware)  # Register
remove_middleware(my_middleware)  # Unregister
get_middlewares()  # List all registered
reset_middlewares()  # Clear all

---

Lifecycle Hooks

Subscribe to SDK events for logging, metrics, or custom behavior:

from swaplm import ChatRequest, ChatResponse, chat, on, off, reset_hooks


def log_request(request: ChatRequest):
    print(f"Sending to {request.model}")


def log_response(response: ChatResponse, request: ChatRequest):
    print(f"Got {response.usage.total_tokens} tokens from {response.provider}")


on("before_request", log_request)
on("after_request", log_response)

response = chat(
    provider="groq",
    model="llama-3.3-70b-versatile",
    messages=[{"role": "user", "content": "Hi"}],
)

off("before_request", log_request)  # Unregister
reset_hooks()  # Clear all

Available Events

Event	Callback Signature	When
before_request	(request: ChatRequest)	Before the HTTP request is sent.
after_request	(response: ChatResponse, request: ChatRequest)	After a successful response.
before_retry	(url, status_or_exc, attempt)	Before a retry attempt.
after_retry	(url, status_or_exc, attempt)	After a retry attempt.
before_stream	(request: ChatRequest)	Before streaming begins.
after_stream	(request: ChatRequest)	After streaming ends.
on_error	(request: ChatRequest, error: SwapLMError)	When an error occurs.

---

Provider & Model Discovery

Query available providers and models programmatically:

import swaplm

# List all providers
for p in swaplm.providers():
    print(f"{p.id}: {p.name} ({p.protocol})")

# Get a specific provider
groq = swaplm.provider("groq")
print(groq.base_url)  # => "https://api.groq.com/openai/v1"

# List all models across all providers
all_models = swaplm.models()
for provider_info, model_info in all_models:
    print(f"{provider_info.id}/{model_info.id}")

# Get a specific model's metadata
provider_info, model_info = swaplm.model("groq/llama-3.3-70b-versatile")
print(model_info.context_window)  # => 128000
print(model_info.supports_tool_calling)  # => True

---

Error Handling

SwapLM maps all provider-specific errors into a unified exception hierarchy:

from swaplm import chat
from swaplm.exceptions import (
    SwapLMError,  # Base — catch everything
    ProviderError,  # Generic provider error
    AuthenticationError,  # 401 / invalid API key
    RateLimitError,  # 429 / rate limited
    TimeoutError,  # Request timed out
    InvalidModelError,  # Model not found (404)
    AmbiguousModelError,  # Alias matches multiple providers
    InvalidProviderError,  # Unknown provider slug
    ConfigurationError,  # SDK configuration issue
    ModelCapabilityError,  # Unsupported operation for model
    ToolCallError,  # Tool/function calling error
)

try:
    response = chat(
        provider="groq",
        model="nonexistent-model",
        messages=[{"role": "user", "content": "Hi"}],
        api_key="invalid",
    )
except AuthenticationError as e:
    print(f"Auth failed: {e}")
    print(f"Provider: {e.provider}")
    print(f"Status: {e.status_code}")
except InvalidModelError as e:
    print(f"Model not found: {e.model}")
except SwapLMError as e:
    print(f"SDK error: {e}")

---

Tool Types Reference

Message

from swaplm import Message

msg = Message(
    role="user",  # "system" | "user" | "assistant" | "tool"
    content="Hello!",  # str | list[dict] | None
    name="optional_name",  # str | None
    tool_calls=[...],  # list[ToolCall] | None
    tool_call_id="call_abc",  # str | None (for role="tool")
)

Tool

from swaplm import Tool

tool = Tool(
    function={
        "name": "my_function",
        "description": "Does something useful",
        "parameters": {
            "type": "object",
            "properties": {
                "query": {"type": "string"},
            },
            "required": ["query"],
        },
    }
)

ChatRequest

from swaplm import ChatRequest, Message

request = ChatRequest(
    model="llama-3.3-70b-versatile",
    messages=[Message(role="user", content="Hi")],
    provider="groq",  # explicit provider (recommended)
    stream=False,
    max_tokens=1024,
    temperature=0.7,
    top_p=0.9,
    stop=["\n\n"],
    seed=42,
    response_format={"type": "json_object"},
    tools=[...],
    tool_choice="auto",
    api_key="optional-explicit-key",
    base_url="https://custom-endpoint.com/v1",
    timeout=30.0,
    retries=3,
    extra_headers={"X-Custom": "value"},
    provider_options={"thinking": {"type": "enabled"}},
)

# Computed properties
request.provider_id  # => "groq"
request.model_id  # => "llama-3.3-70b-versatile"

ModelInfo

from swaplm import model

provider_info, model_info = swaplm.model("groq/llama-3.3-70b-versatile")

model_info.id  # => "llama-3.3-70b-versatile"
model_info.display_name  # => "LLaMA 3.3 70B"
model_info.context_window  # => 128000
model_info.max_tokens  # => 32768
model_info.supports_streaming  # => True
model_info.supports_tool_calling  # => True
model_info.supports_json_mode  # => True
model_info.supports_thinking  # => True
model_info.supports_seed  # => False
model_info.supports_vision  # => False
model_info.default_temperature  # => None
model_info.default_max_tokens  # => None

---

Architecture

chat() / achat()
    │
    ├─ Middleware Pipeline (request)
    │
    ├─ Router ─────────── Resolves "provider/model" string
    │     │
    │     ├─ ProviderRegistry ── Auto-discovers providers
    │     └─ Provider ────────── Metadata + models.json
    │
    ├─ AuthManager ────── Resolves API key (explicit → env var)
    │
    ├─ Protocol ───────── Translates to provider-native format
    │     │
    │     ├─ OpenAIProtocol ─── OpenAI, Groq, NVIDIA, etc.
    │     ├─ AnthropicProtocol  Anthropic Claude
    │     └─ GoogleProtocol ─── Google Gemini
    │
    ├─ Transport ──────── Executes HTTP (httpx / custom)
    │
    ├─ Protocol ───────── Parses response back to unified format
    │
    ├─ Middleware Pipeline (response)
    │
    └─ ChatResponse / StreamResponse

---

Examples

Example	Description
01_basic_chat.py	Synchronous chat with Groq
02_async_chat.py	Async chat with achat()
03_streaming.py	Real-time token streaming
04_tool_calling.py	Function / tool calling
05_structured_output.py	JSON structured output
06_multi_provider.py	Switching between providers
07_middleware.py	Request/response middleware
08_hooks.py	Lifecycle event hooks
09_custom_transport.py	Custom transport injection
10_discovery.py	Provider & model discovery

---

Roadmap

Phase	Milestone	Status
1	Project foundation & repository structure	✅
2	Provider architecture & protocol system	✅
3	First production provider (Groq) & HTTP execution	✅
4	Core SDK infrastructure (routing, discovery, config)	✅
5	Multi-protocol validation (Anthropic & Google Gemini)	✅
6	SDK runtime & developer experience (async, retries, middleware, hooks)	✅
7	OpenAI-compatible provider expansion (16 providers)	✅
8	Enterprise resilience (fallbacks, circuit breakers)	🔜
9	Documentation site & v1.0 release	Planned

---

Contributing

git clone https://github.com/krishcodes07/swaplm.git
cd swaplm
pip install -e ".[dev]"

# Lint & format
ruff check .
ruff format .

# Run tests
pytest

---

License

SwapLM is released under the MIT License.