Railtracks

Own the AI!
Assemble a custom agent harness in plain Python. The loop, the tools, the context, the controls, the record. All yours.

What is Railtracks?

Railtracks is a Python agent framework for building your own agents and/or harness. Every piece is an ordinary Python object you assemble yourself: a tool-calling loop, a tool surface of functions, sub-agents and MCP servers, context management, permission and budget controls, and a replayable record of every run. No YAML, no DSL, no black-box runtime.

import railtracks as rt


# Define a tool (just a function!)
def get_weather(location: str) -> str:
    """Get the current weather for a location."""
    return f"It's sunny in {location}!"


# Create an agent with tools
agent = rt.agent_node(
    "Weather Assistant",
    # Alternatively, use @rt.function_node at def time
    tool_nodes=[rt.function_node(get_weather)],
    llm=rt.llm.OpenAILLM("gpt-5.4-mini"),
    system_message="You help users with weather information.",
)

# Run it
flow = rt.Flow(name="Weather Flow", entry_point=agent)
result = flow.invoke("What's the weather in Paris?")
# or `await flow.ainvoke("What's the weather in Paris?")` in an async context
print(result.text)  # "Based on the current data, it's sunny in Paris!"

Execution order, branching, and looping are expressed using standard Python control flow.

What is an agent harness?

Everything around the model call. The model brings judgment. The harness brings the loop that keeps calling it, the tools it can reach, what lands in its context, the limits on what it may do, and the record of what it did. Change your model tomorrow and the harness is what you still own.

Five parts. Take the ones your problem needs, wire them together in plain Python, leave the rest out.

Part	What it decides	Railtracks primitives
Loop	When the agent keeps going, and when it's done	rt.agent_node runs the tool-calling loop; rt.Flow and rt.call drive multi-step work
Tool surface	What the agent can actually do	rt.function_node, rt.ToolManifest for agents-as-tools, rt.connect_mcp for MCP servers
Context	What the model sees on this turn	system_message, rt.context, todo and key-value memory toolsets, retrieval
Controls	What it's allowed to do, and how much of it	MaxCalls, Timeout, Retry, Lock, human-in-the-loop verifiers, guardrails
Record	What happened, and whether you can replay it	Session state, railtracks viz, rt.evaluations.evaluate

The shapes this usually takes:

• Coding harness. Read, edit, and shell tools, a todo list that survives across turns, human approval on anything that touches the working tree. Runnable in examples/harness/coding_harness.py.
• Research harness. Search and fetch, retrieval over what it has gathered, memory for findings, a structured-output pass to force the report into a schema.
• Operations harness. A few high-consequence tools, each behind a real approver, with an audit trail you can hand to someone else.

Start with the Agent Harness guide, or run the harness examples as they are: a read-only harness in under 80 lines, and a coding harness whose file writes and shell commands each stop for your approval.

Why Railtracks?

Pure Python # Write agents like regular functions @rt.function_node def my_tool(text: str) -> str: return process(text) No YAML, no DSLs, no magic strings Compatible with standard debuggers Full IDE autocomplete and type checking	Tool-First Architecture # Any function becomes a tool agent = rt.agent_node("Assistant", tool_nodes=[my_tool, api_call]) Automatic function-to-tool conversion Seamless API and database integration MCP protocol support
Familiar Interface # Native Async support result = await rt.call(agent, query) Standardized call interface, consistent with asyncio patterns Built-in validation, error handling, and retries Automatic parallelization management	Built-in Observability Railtracks includes a visualizer for inspecting agent runs and evaluations in real-time, run completely locally with no signups required. See the Observability documentation for setup and usage.

Quick Start

Installation

pip install 'railtracks[visual]'

Set your API key

Railtracks loads a local .env file on import, save your provider keys there:

echo "OPENAI_API_KEY=sk-..." >> .env

Your First Agent

import railtracks as rt


# 1. Create tools (just functions with decorators!)
@rt.function_node
def count_characters(text: str, character: str) -> int:
    """Count occurrences of a character in text."""
    return text.count(character)


@rt.function_node
def word_count(text: str) -> int:
    """Count words in text."""
    return len(text.split())


# 2. Build an agent with tools
text_analyzer = rt.agent_node(
    "Text Analyzer",
    tool_nodes=[count_characters, word_count],
    llm=rt.llm.OpenAILLM("gpt-5.4-mini"),
    system_message="You analyze text using the available tools.",
)

# 3. Use it to solve the classic "How many r's in strawberry?" problem
text_flow = rt.Flow(name="Text Analysis Flow", entry_point=text_analyzer)

result = text_flow.invoke("How many 'r's are in 'strawberry'?")
print(result.text)

LLM Support

Railtracks integrates with major model providers through a unified interface:

# OpenAI
rt.llm.OpenAILLM("gpt-5.4-mini")

# Anthropic
rt.llm.AnthropicLLM("claude-sonnet-5")

# Local models
rt.llm.OllamaLLM("llama3")

Works with OpenAI, Anthropic, Google, Azure, and more. See the full provider list.

Contributing

Railtracks is developed in the open. Contributions, bug reports, and feature requests are welcome via GitHub Issues.

---

_{Licensed under MIT · Made by the Railtracks team}