Project Structure

Klisk projects follow a simple, convention-based structure that makes it easy to organize agents and tools.

Default Layout

When you run klisk create my-agent, you get this structure:

my-agent/
├── klisk.config.yaml    # Project configuration
├── .env                 # API keys and secrets (gitignored)
└── src/
    ├── main.py          # Agent definitions
    └── tools/
        ├── __init__.py
        └── example.py   # Tool implementations

This structure is a convention, not a requirement. You can organize your code however you like as long as the entry point is specified in klisk.config.yaml.

Configuration File

klisk.config.yaml

The root configuration file defines project metadata and entry point:

klisk.config.yaml

entry: src/main.py
name: "MyAgent"
studio:
  port: 3000
api:
  port: 8000

entry

string

required

Path to the Python file that defines your agents (relative to project root)

name

string

default:"MyAgent"

Display name for the project (shown in Studio header)

studio.port

number

default:3000

Port for the Studio web interface (currently unused, defaults to 8000)

api.port

number

default:8000

Port for the dev server and API endpoints

If klisk.config.yaml is missing, Klisk uses default values and assumes src/main.py as the entry point.

Entry Point

src/main.py

The entry file imports and defines your agents:

src/main.py

from klisk import define_agent, get_tools

agent = define_agent(
    name="Assistant",
    instructions="You are a helpful assistant. Use the tools available to help the user.",
    model="gpt-5.2",
    tools=get_tools("greet"),
)

Klisk automatically discovers all define_agent() calls in your entry file and registers them. You can define multiple agents in a single file or split them across modules.

Module Imports

You can organize agents across multiple files:

src/main.py

from klisk import define_agent, get_tools
from .specialized_agents import researcher, writer

assistant = define_agent(
    name="Assistant",
    model="gpt-5.2",
    tools=get_tools("greet"),
)

src/specialized_agents.py

from klisk import define_agent, get_tools

researcher = define_agent(
    name="Researcher",
    model="gpt-5.2",
    tools=get_tools("web_search"),
    builtin_tools=["web_search"],
)

writer = define_agent(
    name="Writer",
    model="anthropic/claude-sonnet-4-5",
    instructions="You are a professional writer.",
)

All agent definitions must be imported into the entry file (directly or indirectly) to be discovered by Klisk.

Tools Directory

src/tools/

Tools are Python functions decorated with @tool. Klisk discovers them automatically from any .py file in the project.

src/tools/example.py

from klisk import tool

@tool
async def greet(name: str) -> str:
    """Greet someone by name."""
    return f"Hello, {name}! How can I help you today?"

Tool Discovery

Klisk scans your project for @tool decorators:

Walks all .py files in the project directory
Parses each file with AST to find decorated functions
Registers tools by function name
Extracts docstrings as descriptions and type hints as parameters

# src/tools/greetings.py
from klisk import tool

@tool
async def greet(name: str) -> str:
    """Greet someone by name."""
    return f"Hello, {name}!"

@tool
async def farewell(name: str) -> str:
    """Say goodbye to someone."""
    return f"Goodbye, {name}!"

Tool Naming

Tool names are derived from function names:

Function greet → tool "greet"
Function fetch_weather → tool "fetch_weather"

Reference them in agents using get_tools():

agent = define_agent(
    name="Assistant",
    model="gpt-5.2",
    tools=get_tools("greet", "fetch_weather"),
)

Tool names must be unique across your project. If you define two tools with the same name, the last one discovered will overwrite earlier definitions.

Environment Variables

.env

Store API keys and secrets in a .env file at the project root:

.env

OPENAI_API_KEY=sk-your-key-here

# Uncomment and fill in the API key for your chosen provider:
# ANTHROPIC_API_KEY=sk-ant-your-key-here
# GEMINI_API_KEY=your-key-here
# MISTRAL_API_KEY=your-key-here

Klisk loads this file automatically when you run klisk dev or klisk run:

# From dev.py
from dotenv import load_dotenv
from klisk.core.config import ProjectConfig

project_path = resolve_project(name_or_path)
load_dotenv(project_path / ".env")

config = ProjectConfig.load(project_path)

Never commit .env to version control. The default .gitignore includes .env to prevent accidental exposure of secrets.

Workspace Directory

~/.klisk/projects/

Klisk uses a global workspace directory to store projects:

~/.klisk/
└── projects/
    ├── my-agent/
    │   ├── klisk.config.yaml
    │   ├── .env
    │   └── src/
    ├── researcher/
    └── writer/

Run klisk dev without arguments to work with all projects in this directory simultaneously.

Project Resolution

When you run klisk dev my-agent, Klisk resolves the project path:

If my-agent is an absolute or relative path, use it directly
Otherwise, look for ~/.klisk/projects/my-agent/
If not found, raise an error

# From paths.py (simplified)
def resolve_project(name_or_path: str) -> Path:
    path = Path(name_or_path)
    if path.is_absolute() or path.exists():
        return path.resolve()
    
    workspace_path = PROJECTS_DIR / name_or_path
    if workspace_path.exists():
        return workspace_path
    
    raise ValueError(f"Project not found: {name_or_path}")

Ignored Paths

Klisk ignores certain directories when scanning for tools and watching for changes:

.git/, .venv/, __pycache__/, node_modules/
Any directory starting with . (hidden directories)
Files outside .py, .yaml, .yml extensions (for watching)

# From watcher.py
def _is_relevant(path: str) -> bool:
    p = Path(path)
    parts = p.parts
    if any(part.startswith(".") or part == "__pycache__" 
           or part == "node_modules" or part == ".venv" 
           for part in parts):
        return False
    return p.suffix in {".py", ".yaml", ".yml"}

Multi-Project Workspace

You can work with multiple projects in a workspace:

# Create multiple projects
klisk create researcher
klisk create writer
klisk create coordinator

# Launch Studio for all projects
klisk dev

# The Studio lists all agents from all projects
# Each agent is prefixed with its project name:
# - researcher/Researcher
# - writer/Writer
# - coordinator/Coordinator

In workspace mode, agent and tool names are prefixed with the project name to avoid collisions (e.g., researcher/search vs writer/search).

Custom Layouts

You can use any structure as long as you update klisk.config.yaml:

entry: agents/main_agent.py
name: "CustomLayout"

Klisk will discover tools from anywhere in the project directory, so you can organize them however makes sense:

custom-project/
├── klisk.config.yaml
├── agents/
│   └── main_agent.py
├── tooling/
│   ├── api_calls.py
│   └── data_processing.py
└── utils/
    └── helpers.py

Best Practices

Keep tools focused and single-purpose

Each tool should do one thing well. This makes them easier to test, compose, and understand.

Use descriptive docstrings

The docstring becomes the tool description shown to the LLM. Make it clear and actionable.

Organize tools by domain

Group related tools in subdirectories: src/tools/web/, src/tools/data/, etc.

Avoid circular imports

Don’t import agents from tool files or vice versa. Keep dependencies unidirectional.

Use .env for all secrets

Never hardcode API keys or credentials. Always use environment variables.

File Watching Behavior

The dev server watches your project and reloads when files change:

# From app.py
async def _on_file_change() -> None:
    global _snapshot
    try:
        if _workspace_mode:
            _snapshot = discover_all_projects()
        else:
            _snapshot = discover_project(_project_path)
    except Exception as e:
        logger.exception("Failed to reload project")
        _snapshot = ProjectSnapshot()
        _snapshot.config = {"error": str(e)}

    # Notify all connected Studio clients
    data = json.dumps({"type": "reload", "snapshot": _snapshot.to_dict()})
    for ws in _reload_clients:
        await ws.send_text(data)

This enables hot reload without restarting the server.

Get Started

Core Concepts

Development

Deployment

Project Structure

Default Layout

Configuration File

klisk.config.yaml

Entry Point

src/main.py

Module Imports

Tools Directory

src/tools/

Tool Discovery

Tool Naming

Environment Variables

.env

Workspace Directory

~/.klisk/projects/

Project Resolution

Ignored Paths

Multi-Project Workspace

Custom Layouts

Best Practices

File Watching Behavior

Build docs developers (and LLMs) love

Get Started

Core Concepts

Development

Deployment

Documentation Index

​Default Layout

​Configuration File

​klisk.config.yaml

​Entry Point

​src/main.py

​Module Imports

​Tools Directory

​src/tools/

​Tool Discovery

​Tool Naming

​Environment Variables

​.env

​Workspace Directory

​~/.klisk/projects/

​Project Resolution

​Ignored Paths

​Multi-Project Workspace

​Custom Layouts

​Best Practices

​File Watching Behavior

Build docs developers (and LLMs) love

Default Layout

Configuration File

klisk.config.yaml

Entry Point

src/main.py

Module Imports

Tools Directory

src/tools/

Tool Discovery

Tool Naming

Environment Variables

.env

Workspace Directory

~/.klisk/projects/

Project Resolution

Ignored Paths

Multi-Project Workspace

Custom Layouts

Best Practices

File Watching Behavior