AgentCore is the reasoning engine at the heart of CLIver. It implements the Re-Act loop — send messages to an LLM, execute tools, feed results back, and continue until a final response is ready.

The Re-Act Loop

1. Receive user input
2. Build messages (system prompt + conversation history + user input)
3. Send to LLM
4. If the LLM returns a tool call → execute tools → go to step 3
5. If the LLM returns text → done, return the response

The loop runs up to 50 iterations by default, with a safety limit of 5 consecutive tool errors before stopping.

Quick Example

import asyncio

from cliver_core import AgentCore, ModelBindings, presets
from cliver_core.tools import discover_builtin_tools


async def main():
    agent = AgentCore(
        ModelBindings(llm=presets.deepseek(api_key="...")),
        builtin_tools=discover_builtin_tools(),
    )
    response = await agent.chat("List the files in this directory.")
    print(response.message.content)
    await agent.aclose()


asyncio.run(main())

Key Properties

• Embeddable. AgentCore has zero CLI dependencies — it takes Python objects and returns message objects.
• Stateless instance, stateful calls. Conversation context is passed per call (conversation=), so one instance is safe for concurrent chats.
• Bindings, not config files. Every model purpose — chat, image, video, TTS, ASR, embed, rerank — lives in ModelBindings.
• Lazy call-time resolution. Extra tools, MCP servers, tool filters, and skill dispatch are resolved per chat()/stream() call.
• Observable. Pass on_event= to receive inference and tool lifecycle events, including tool.start, tool.end, tool.error, and tool.permission_denied.

Calling the Agent

chat() returns the final CLIverResponse; stream() yields CLIverMessageChunk objects as content arrives. Both accept the same per-call options:

Parameter	Description
user_input	The user’s message
system_prompt	Additional system content for this call
conversation	Previous messages for multi-turn context
media	Images/audio/video attached to the user message
extra_tools	Additional tools for this call only
mcp_servers	MCP server names to advertise for this call
tool_filter	Predicate that narrows the advertised tool set
options	Wire request options; call site overrides binding defaults
max_iterations	Max Re-Act iterations (default 50)
execution_policy	Ambient tool-execution policy for this call
skill_dispatch	Advertise the skill catalogue and dispatch tools (default on)

When finished, call await agent.aclose() to close providers the agent built.