Synapse — The AI Application Runtime

pip install synapse-runtime

Not a framework. Not a wrapper. The missing AI layer for your Python stack.

30-Second Demo

from synapse import agent, tool, workflow, step

@tool
def search_web(query: str) -> list[str]:
    """Search the web. Returns top results."""
    ...  # your implementation

@agent(model='gpt-4o', fallback='ollama/llama3')
async def researcher(topic: str) -> str: ...

@agent(model='claude-3-5-sonnet')
async def writer(content: str) -> str: ...

# Chain agents with >>
pipeline = researcher >> writer
result   = await pipeline.run('AI application runtimes')

That's it. No subclassing. No boilerplate. No JSON schema by hand.

---

The Four Primitives (v0.1)

@tool — Auto-generated schemas

@tool
def send_email(to: str, subject: str, body: str = '') -> bool:
    """Send an email. Returns True on success."""
    ...

Synapse reads the type annotations and docstring. The JSON schema is generated automatically — no BaseTool, no {type: "function", function: {…}} by hand.

Pass tools explicitly to the agents that need them:

@agent(model='gpt-4o', tools=[search_web])   # only this agent can call search_web
async def researcher(topic: str) -> str: ...

@agent(model='gpt-4o')                        # clean — no tools injected
async def writer(draft: str) -> str: ...

Each agent only receives the schemas you declare. No global bleed between agents.

@agent — Any model, same API

@agent(model='gpt-4o')                  # OpenAI
@agent(model='claude-3-5-sonnet')       # Anthropic
@agent(model='gemini/gemini-pro')       # Google
@agent(model='ollama/llama3')           # Local (free)

Powered by litellm — 100+ providers, one decorator.

Add guardrails:

@agent(
    model='gpt-4o',
    fallback='ollama/llama3',
    must_not=['PII', 'speculation'],
    token_budget=2000,
)
async def safe_agent(prompt: str) -> str: ...

>> — Readable pipelines

pipeline = researcher >> writer >> reviewer
result   = await pipeline.run(topic='climate change')

LangGraph needs nodes and edges. Synapse uses >>.

@workflow + step() — Composable pipelines

@workflow
async def research_pipeline(topic: str) -> str:
    data  = await step(researcher, topic)
    draft = await step(writer, data)
    return await step(reviewer, draft)

---

Zero-Cost Testing

from synapse import MockLLM

def test_my_agent():
    with MockLLM("expected response") as mock:
        result = await my_agent("prompt")
    assert result == "expected response"
    assert mock.call_count == 1

MockLLM patches litellm. No API calls. No tokens spent.

---

Synapse vs the alternatives

Feature	Synapse	LangChain	OpenAI SDK	LangGraph
Any LLM provider	✅	✅	❌ OpenAI only	✅
Schema from annotations	✅	❌ manual	❌ manual	❌ manual
>> chaining	✅	❌ verbose	❌	❌
Built-in guardrails	✅	⚠️ external	❌	❌
Readable 3-line agent	✅	❌ subclass	❌	❌
MockLLM for tests	✅	❌	❌	❌

---

Roadmap

Version	Name	Ships
v0.1	The Tiny Core	@agent, @tool, >>, MockLLM ← you are here
v0.3	Useful Projects	Memory API, streaming, prompt templates, CLI
v0.5	AI Runtime	Workflow engine, checkpointing, observability, guardrails
v0.8	Synapse Identity	Prompt debugger, multi-agent graphs, human-in-the-loop
v1.0	Production	Multi-tenancy, audit trail, intent routing

---

Install

pip install synapse-runtime

# For local LLM (free, no API key needed):
# Install Ollama → https://ollama.ai
ollama pull llama3

MIT License · Built in public · Issues welcome