API Reference

Quick reference for all public classes, functions, and types.

---

Agent

from exagent import Agent

Base class for all agents.

Constructor pattern

class MyAgent(Agent):
    def __init__(self):
        self.system_description = "..."   # optional — sets the system prompt
        self.set_model(provider, model)   # required
        self.add_tool(my_tool)            # optional, repeatable
        self.load_system_skill("path")    # optional, repeatable
        super().__init__()                # always last

Methods

Method	Signature	Description
set_model	(provider: str, model: str, **kwargs) → Model	Configure the LLM. Accepted providers: "openai", "anthropic".
add_tool	(tool: Tool \| Callable) → Tool	Register a tool. Accepts a Tool instance or a plain function.
add_tools	(tools: list) → list[Tool]	Register multiple tools at once.
load_system_skill	(path: str) → dict	Load a skill from a markdown file and register it.
load_system_skills	(paths: list[str]) → list[dict]	Load multiple skills at once.
run	(prompt, max_iterations=10, on_tool_call=None, on_iteration=None) → str	Drive the agent loop (blocking) and return the final text.
stream	(prompt, max_iterations=10, on_tool_call=None, on_iteration=None) → Iterator[dict]	Drive the agent loop (streaming) and yield events.

Attributes

Attribute	Type	Description
chat_history	list[dict]	Full conversation history. chat_history[0] is always the system turn.
system_prompt	str	The assembled system prompt (base description + injected skills).
tools	list[Tool]	Registered tools.
skills	list[dict]	Loaded skills. Each dict has name, description, content.
model	Model \| None	The configured model. None until set_model() is called.

---

OrchestratorAgent

from exagent.multi_agent.orchestrator import OrchestratorAgent

Extends Agent. Routes tasks to registered specialist agents. All Agent methods are inherited.

Class attributes

Attribute	Type	Default	Description
parallel	bool	False	Allow the model to call multiple agents in one turn. Default is serial. OpenAI only.

Additional method

Method	Signature	Description
add_agent	(agent: Agent, name: str, description: str \| None = None) → None	Register a specialist as a callable tool. description falls back to agent.system_description.

Constructor pattern

class MyOrchestrator(OrchestratorAgent):
    parallel = False  # default — serial calls (recommended)

    def __init__(self):
        self.set_model("openai", "gpt-4.1-mini")
        self.add_agent(SpecialistA(), name="specialist_a", description="...")
        self.add_agent(SpecialistB(), name="specialist_b", description="...")
        super().__init__()

---

@tool decorator

from exagent import tool

Turns a plain Python function into a Tool. Schema is inferred from type hints; description from the docstring.

@tool
def my_function(x: str, y: int = 0) -> str:
    """Does something."""
    ...

# Override name or description
@tool(name="custom_name", description="Custom description.")
def my_function(...):
    ...

---

Tool

from exagent import Tool

A callable tool the model can invoke.

Constructor

Tool(
    name: str,
    description: str,
    parameters: dict,   # JSON Schema object
    handler: Callable,
)

Methods

Method	Description
tool.run(arguments: dict)	Invoke the handler with a dict of arguments.
tool.to_anthropic()	Return the Anthropic-format tool definition dict.
tool.to_openai()	Return the OpenAI Responses API-format tool definition dict.

---

shell()

from exagent import shell

Start an interactive terminal session with an agent.

shell(
    agent: Agent,
    *,
    stream: bool = True,
    prompt: str = "> ",
) -> None

Command	Effect
/help	Show available commands
/clear	Reset conversation history (keeps system prompt)
/exit / /quit	Exit the shell
Ctrl-C / Ctrl-D	Exit cleanly

---

ToolCall

from exagent import ToolCall

Represents a tool-use request emitted by the model.

Attribute	Type	Description
id	str	Unique ID for this call.
name	str	Name of the tool being called.
input	dict	Arguments the model passed.

---

ProviderResponse

from exagent import ProviderResponse

The structured result from a model turn.

Attribute	Type	Description
text	str	Text content of the response.
tool_calls	list[ToolCall]	Tool calls requested this turn. Empty if the model stopped.
assistant_message	dict \| None	Canonical assistant turn appended to history.
stop_reason	str \| None	Why the model stopped ("end_turn", "tool_use", etc.)

---

Stream event shapes

All events yielded by agent.stream():

{"type": "text_delta",  "text": str}
{"type": "tool_call",   "tool_call": ToolCall}
{"type": "tool_result", "id": str, "name": str, "content": str, "is_error": bool}
{"type": "done",        "text": str}

---

Utility functions

from exagent import load_skill, load_file_as_string

load_skill(path: str) -> dict          # parse a skill markdown file
load_file_as_string(path: str) -> str  # read a file as a plain string