KHarness

A lightweight Python framework for building LLM agents with tool use, conversation memory, and multi-round research pipelines. Supports local inference via llama.cpp and Ollama.

---

Installation

pip install kharness

Optional dependencies (install what you need):

pip install langchain-core          # Required for tool decorators
pip install ddgs beautifulsoup4     # Web tools
pip install yahoors                 # Stock tools

---

Providers

Providers wrap a local inference backend and expose a unified chat() interface.

Ollama

from kharness.providers import Ollama

provider = Ollama("qwen2.5:7b")             # model is required

response = provider.chat(
    [{"role": "user", "content": "Hello!"}]
)

Parameter	Default	Description
model	—	Model name (required)
server_url	"http://localhost"	Server host
port	"11434"	Server port
max_iterations	10	Max tool-call loop iterations

LlamaCpp

Assumes a llama-server process is already running. Connect to it:

from kharness.providers import LlamaCpp

provider = LlamaCpp()                       # defaults: 127.0.0.1:8080

if provider.check_health():
    response = provider.chat(
        [{"role": "user", "content": "Hello!"}]
    )

Parameter	Default	Description
server_url	"http://127.0.0.1"	Server host
port	"8080"	Server port
max_iterations	5	Max tool-call loop iterations
strip_tools_after	3	Force text response after N tool rounds

---

Agent

Agent wraps a provider with optional tools and memory, exposing run() for chat and research() for multi-round deep dives.

Basic chat

from kharness.providers import Ollama
from kharness.agent import Agent

provider = Ollama("qwen2.5:7b")

agent = Agent(provider)
response = agent.run("What is the capital of France?")
print(response)

With a system prompt file

agent = Agent(provider, soul_md_path="persona.md")

With tools

Pass a dict or list of dicts mapping tool names to LangChain @tool objects:

from kharness.tools import WEB_TOOL_MAP

agent = Agent(provider, tool_map=WEB_TOOL_MAP)
response = agent.run("What happened in the news today?")

With memory

Attach a Memory object to maintain conversation history across turns:

from kharness.memory import Memory

memory = Memory(max_turns=20)
agent = Agent(provider, memory=memory)

agent.run("My name is Alice.")
agent.run("What's my name?")   # remembers "Alice"

---

Memory

Memory manages a sliding window of conversation history. It stores messages as {"role": ..., "content": ...} dicts — the format all providers expect.

from kharness.memory import Memory

memory = Memory(max_turns=10)   # keeps last 10 user/assistant pairs
memory.add("user", "Hello")
memory.add("assistant", "Hi there!")

history = memory.get_history()  # list of message dicts
memory.clear()

Parameter	Default	Description
max_turns	20	Max user/assistant pairs to retain

---

Research pipeline

agent.research() runs a structured multi-round research loop:

1. Plan — generates search queries from the question
2. Gather — runs web searches for each query
3. Analyze — synthesizes findings and decides if more searching is needed
4. Repeat — loops up to max_rounds times if gaps remain
5. Report — writes a final comprehensive answer

from kharness.providers import Ollama
from kharness.tools import WEB_TOOL_MAP
from kharness.agent import Agent

provider = Ollama("qwen2.5:14b")

agent = Agent(provider, tool_map=WEB_TOOL_MAP)

report = agent.research(
    "What are the latest developments in fusion energy?",
    max_rounds=3,
    planning_tokens=512,
    analysis_tokens=2048,
    report_tokens=8000,
)
print(report)

Parameter	Default	Description
max_rounds	3	Maximum search/analyze iterations
planning_tokens	1024	Token budget for query planning
analysis_tokens	2048	Token budget for per-round analysis
report_tokens	8000	Token budget for final report
debug	True	Log phase progress

---

Tools

Tools are standard LangChain @tool functions. Convenience maps are exported for easy use with Agent.

Web tools

from kharness.tools import WEB_TOOL_MAP

Tool	Description
web_search	DuckDuckGo text search, returns top snippets
fetch_page	Fetches a URL and extracts readable text
news_search	DuckDuckGo news search with dates
image_search	Returns direct image URLs
wikipedia_summary	Wikimedia REST API summary
search_and_fetch	Combines search + full page fetch
search_subreddit	Fetches top posts from a subreddit

Stock tools

from kharness.tools import STOCK_TOOL_MAP

Tool	Description
get_candles	OHLCV candlestick data
get_options	Options chain (calls and puts)
get_income_statements	Income statement financials
get_balance_sheet	Balance sheet financials
get_cash_flow	Cash flow statement

Custom tools

Any LangChain @tool works:

from langchain_core.tools import tool

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

agent = Agent(provider, tool_map={"add": add})

---

Logging

KHarness uses Python's standard logging module. To see runtime output:

import logging
logging.basicConfig(level=logging.INFO)

Use logging.DEBUG to see individual tool calls.

---

Requirements

• Python >= 3.12
• A running Ollama daemon or llama-server process