sAI SDK · version 0.2.0 · macOS

sAI's agents.
Inside your app.

Bring your own model. Get a whole team of coding agents — with tools, approvals and a sandbox — in a few lines of Python.

Quick start

The SDK comes with sAI — install and activate it as usual, then:

from sai_sdk import Agent, LLM

agent = Agent(LLM.openai("gpt-4o-mini", api_key="sk-..."),
              folder="~/my-project", allow_edits=True)

for event in agent.run("fix the failing test"):
    print(event)          # every step as it happens; the last one is the Result

result = agent.ask("what does utils.py do?")   # or just the final answer
print(result.answer)

Your model

Use any model with an OpenAI-compatible API. Agents need tool calling to read files, edit and run commands; without it they can still answer questions.

LLM("gpt-4o-mini", api_key=...)                       # OpenAI
LLM("deepseek-chat", base_url="https://api.deepseek.com/v1", api_key=...)
LLM.openrouter("qwen/qwen3-coder", api_key=...)
LLM.groq("openai/gpt-oss-120b", api_key=...)          # free tier
LLM.gemini("gemini-3.5-flash", api_key=...)           # free tier
LLM.ollama("qwen2.5-coder:14b")                       # on your computer, no key
LLM.custom(my_function)                               # anything else

Your key is used for calls only and never stored.
Rate limits and temporary errors are retried automatically, waiting as long as the provider asks.
LLM.custom(fn): fn(messages, tools) receives OpenAI-format messages and returns a string, or {"content": ..., "tool_calls": [...]}.
Smaller models pick agents less reliably — pass agent="Coder" to choose one yourself.

Agent

Option	What it does
`llm`	Your model (required).
`folder`	The project folder. Agents can't reach outside it.
`tools`	Your own tools — see Your tools.
`approve`	Your function that allows or refuses each file edit and risky command.
`allow_edits`	Without `approve`: apply edits automatically. Risky commands are refused.
`agent`	Always use one agent: Coder, Debugger, Researcher, Reviewer, Writer, Planner, Architect, DevOps.
`max_steps`	Tool steps per request (default 12).
`sandbox`	Run agent commands in the macOS sandbox (default on).

run(prompt) streams events · ask(prompt) returns the Result · reset() starts a new conversation (otherwise it continues) · stop() stops the current run.

Events

Event	Means
`Notice`	Which agent took the request.
`Text`	Reply text as it's being written.
`ToolCall`	An agent is using a tool, with its arguments.
`ToolResult`	What the tool returned, and whether it worked.
`AgentMessage`	A complete message; `final` marks the answer.
`Result`	Always last: `answer`, `agent`, `stop_reason`, `declined`, `models`, tokens, `error`.

Model and tool failures never crash your program — they arrive as a Result with stop_reason="error".

Approvals

def approve(request):
    print(request.kind, request.description)   # "edit", "command", ...
    print(request.diff)                        # for edits: exactly what changes
    return ask_my_user()                       # True to allow

agent = Agent(llm, folder="~/my-project", approve=approve)

A refused action isn't retried or worked around — the agent explains what it wanted to do, and result.declined lists it.

Your tools

from sai_sdk import tool

@tool
def ticket(ticket_id: str) -> str:
    "Look up a bug ticket by its id."       # what the model sees
    return my_tracker.get(ticket_id)

agent = Agent(llm, folder="~/my-project", tools=[ticket])
agent.ask("Fix the bug described in ticket BUG-142.")

Typed parameters become the tool's arguments; those without defaults are required. Return text. If your function raises, the agent is told and can react.

Good to know

One run at a time per Python process. Run agents one after another, or in separate processes.
macOS and Python 3.14, like the rest of sAI 0.2.0.
The SDK needs your activated sAI license, on the computer it was issued for.
Your code and prompts go only to the model provider you choose.