picoflow 0.1.9

Simple, flexible AI agent framework with a small DSL and explicit state

PicoFlow — Simple, Flexible AI Agent Framework

Build agents with explicit steps and a small DSL.
LLMs, tools, loops, and branches compose naturally.

---

A Minimal PicoFlow Application

from picoflow import flow, llm, create_agent

LLM_URL = "llm+openai:///gpt-4.1-mini?api_key_env=OPENAI_API_KEY"

@flow
async def mem(ctx):
    return ctx.add_memory("user", ctx.input)

agent = create_agent(
    mem >> llm("Answer in one sentence: {input}", llm_adapter=LLM_URL)
)

print(agent.get_output("What is PicoFlow?", trace=True))

export OPENAI_API_KEY=sk-...
python minimal.py

---

Core Ideas

• Flow = step
  A flow is just a Python function that takes and returns State.

• DSL = pipeline
  Use >> to compose steps into readable execution graphs.

• Agent = runner
  create_agent(flow) gives you run / arun / get_output.

• State = context (Ctx)
  Ctx is an alias of State. It is immutable and explicit.

---

Quick Start (Step by Step)

1. Define Steps with @flow

from picoflow import flow, Ctx

@flow
async def normalize(ctx: Ctx) -> Ctx:
    return ctx.update(input=ctx.input.strip().lower())

---

2. Call LLM as a Step

from picoflow import llm

ask = llm("Answer briefly: {input}")

---

3. Compose with DSL

pipeline = normalize >> ask

---

4. Run with Agent

from picoflow import create_agent

agent = create_agent(pipeline)

state = await agent.arun("Hello WORLD")
print(state.output)

---

DSL in One Minute

Sequential

flow = a >> b >> c

Loop

flow = step.repeat()

or:

flow = repeat(step, until=lambda s: s.done)

Parallel + Merge

flow = fork(a, b) >> merge()

Custom merge:

flow = fork(a, b) >> merge(
    mode=MergeType.CUSTOM,
    reducer=lambda branches, main: branches[0]
)

---

LLM URL

from picoflow.adapters.registry import from_url

adapter = from_url(
    "llm+openai:///gpt-4.1-mini?api_key_env=OPENAI_API_KEY"
)

Then:

flow = llm("Explain: {input}", llm_adapter=adapter)

Custom Adapters

class MyAdapter(LLMAdapter):
    def __call__(self, prompt: str, stream: bool):
        ...

from picoflow.adapters.registry import register

register("myllm", lambda url: MyAdapter(...))

Use:

llm+myllm://host/model?param=value

OpenAI-Compatible DSN Notes

The OpenAI-compatible adapter uses this shape:

llm+openai://<host>/<model>?k=v&...

Examples:

llm+openai:///gpt-4.1-mini?api_key_env=OPENAI_API_KEY
llm+minimax:///MiniMax-Text-01?api_key_env=MINIMAX_API_KEY
llm+openai://ark.cn-beijing.volces.com/doubao-seed-1-8-251228?api_key_env=OPENAI_API_KEY
llm+minimax://api.minimaxi.com/MiniMax-Text-01?api_key_env=MINIMAX_API_KEY
llm+openai://127.0.0.1:8000/Qwen2.5?api_key=none

Supported standard query params include:

• api_key
• api_key_env
• base_url
• base_path
• timeout
• verify
• insecure
• ca_file
• ca_path

All other query params are passed through to the request payload automatically.

Example: Doubao thinking

llm+openai://ark.cn-beijing.volces.com/doubao-seed-1-8-251228?api_key_env=OPENAI_API_KEY&thinking={"type":"enabled"}

Example: OpenAI-compatible reasoning parameter

llm+openai:///gpt-5.1?api_key_env=OPENAI_API_KEY&reasoning_effort=medium

Example: MiniMax via the OpenAI-compatible adapter alias

llm+minimax:///MiniMax-Text-01?api_key_env=MINIMAX_API_KEY

Simple JSON values are parsed automatically:

• JSON objects and arrays become dict/list
• JSON numbers become int/float
• true / false / null become boolean / None
• other values stay as strings

Core fields such as model, messages, and stream are still controlled by PicoFlow and are not taken from the query string.

---

Runtime Options

Tracing

await agent.arun("hi", trace=True)

Timeout

await agent.arun("hi", timeout=10)

Streaming

async def on_chunk(text: str):
    print(text, end="", flush=True)

await agent.arun("stream me", stream_callback=on_chunk)

---

Tools

from picoflow import tool

flow = tool("search", lambda q: {"result": "..."} )

Results:

state.tools["search"]

Troubleshooting: SSL certificate verify failed

Recommended (secure): set your CA bundle

export SSL_CERT_FILE=/path/to/ca.pem
# or
export PICO_CA_FILE=/path/to/ca.pem

Quick debug (insecure): disable verification temporarily

...&insecure=1

or

export PICO_SSL_VERIFY=0

Do not use this in production.

Note: Local Ollama usage (llm+ollama://localhost:11434/...) uses plain HTTP and does not require SSL configuration.

---

📚 Cookbook (Examples)

The cookbook/ directory contains small, focused examples for common patterns.
Each folder is runnable and demonstrates one specific feature or usage style.

cookbook/
├─ minimal-demo        # Smallest runnable PicoFlow example
├─ multiple-crew       # Multi-agent / crew-style collaboration
├─ simple-chat         # Basic chat agent
├─ simple-chat-stream  # Streaming responses from LLM
├─ simple-tool         # Tool calling and tool state
└─ trace               # Tracing and execution visualization

How to run

All examples are standalone scripts:

export OPENAI_API_KEY=sk-...
cd cookbook/simple-chat
python main.py

When to use which example

• Start here → minimal-demo
  Understand Flow, DSL (>>), and Agent execution.

• LLM interaction → simple-chat, simple-chat-stream
  Prompt → response, with and without streaming.

• Tool calling → simple-tool
  How tools are defined, executed, and stored in state.tools.

• Multi-step / multi-agent → multiple-crew
  Composition of multiple flows and roles.

• Debugging & observability → trace
  How tracing hooks and structured events work.

Cookbook examples are intentionally small and close to real usage.
They are recommended as the primary learning path after reading the Minimal Example.

---

License

MIT