jiki 0.0.8

A flexible LLM orchestration framework with tool calling capabilities via MCP protocol

Jiki

Jiki is a flexible LLM orchestration framework with built-in tool calling capabilities.

Overview

Jiki provides a clean interface for building AI assistants that can use tools to solve problems. It orchestrates the interaction between language models and external tools using the MCP (Model Context Protocol).

Features

• Seamless integration with LiteLLM for support of multiple LLM providers
• Tool calling through FastMCP infrastructure (manual config or auto-discovery)
• Flexible MCP client with multiple transport options (stdio, SSE)
• Structured conversation logging for training data generation
• Simple CLI interface for interactive use
• XML-based tool call format for clear model interaction

Quick Start

uv add jiki

from jiki import create_jiki

# Create a pre‑configured orchestrator with sensible defaults
# Option 1: Provide tool schemas manually
# orchestrator = create_jiki(
#     model="anthropic/claude-3-7-sonnet-latest",
#     # Path to a JSON file that describes your tools (see below)
#     tools="tools.json", 
#     mcp_mode="stdio", # Use stdio for a local FastMCP server script
#     mcp_script_path="path/to/your/server.py" 
# )

# Option 2: Auto-discover tools from a running FastMCP server
orchestrator = create_jiki(
    model="anthropic/claude-3-7-sonnet-latest",
    auto_discover_tools=True,  # Discover tools from the server
    mcp_mode="stdio",          # Transport mode ('stdio' or 'sse')
    mcp_script_path="servers/calculator_server.py" # Path to server script (or URL for SSE)
)

# Process a user query (synchronous helper available)
result = orchestrator.process("What is 2 + 2?")
print(result)

CLI Usage

uv run examples/simple_multiturn_cli.py --tools tools.json

Detailed Responses & Tracing

Jiki can return a rich DetailedResponse object that includes the assistant's answer and all tool calls / raw traces generated during the turn. This is useful for debugging, analytics, or offline reinforcement‑learning pipelines.

# Get a structured response with trace metadata
detailed = orchestrator.process_detailed("What is the capital of France?")

print(detailed.result)      # The assistant's final answer
print(detailed.tool_calls)  # List[ToolCall] detailing every tool invocation
print(detailed.traces)      # Raw trace dictionaries for deeper inspection

# Persist traces from the current session
orchestrator.export_traces("interaction_traces/session.jsonl")

Creating Custom Tools

Tools are defined in JSON format and implemented using FastMCP:

[
  {
    "tool_name": "add",
    "description": "Add two numbers",
  "arguments": {
    "a": {"type": "integer", "description": "First number"},
    "b": {"type": "integer", "description": "Second number"}
  }
}

Server implementation (servers/calculator_server.py):

from fastmcp import FastMCP

mcp = FastMCP("Calculator")

@mcp.tool()
def add(a: int, b: int) -> int:
    """Add two numbers"""
    return a + b

if __name__ == "__main__":
    mcp.run()

Requirements

• Python 3.11+
• litellm >= 1.35.0 (or the latest)
• fastmcp >= 2.1.1
• mcp
• tiktoken (optional – enables exact token counting)