Getting started

Quickstart

Build a governed LLM agent in a few lines. Pick a model, register tools, set the rules — the runtime enforces them on every step.

A governed agent

agent.py
import asyncio
from antraft import Antraft, ToolRegistry, tool_from_function
from antraft.models import create_model
 
def search_web(query: str) -> str:
    "Search the web."
    return f"results for {query}"
 
tools = ToolRegistry([tool_from_function(search_web)])
 
runtime = (
    Antraft.agent(
        create_model("anthropic:claude-opus-4-8"),
        tools,
        task="Research Antraft and summarize it.",
        guardrails=True,
    )
    .allow(["search_web"])       # default-deny everything else
    .budget(max_cost_usd=1.0)    # enforce spend as policy
    .build()
)
 
asyncio.run(runtime.run())
Note
The parameter schema for search_web is inferred from its type hints, so the model knows how to call it and the gateway validates the arguments before it runs.

Inspect the result

agent.py
ctx = asyncio.run(runtime.run())
 
print(runtime.agent.final_answer)   # the model's answer
print(ctx.action_history)           # tools that actually ran
print(ctx.tokens_used, ctx.cost_usd)
print(ctx.completed, ctx.killed)

Everything switched on

agent.py
from antraft import Antraft, SemanticMemory
 
runtime = (
    Antraft.agent(
        model, tools,
        task="...",
        guardrails=True,                       # injection / jailbreak / PII / secrets
        memory=SemanticMemory(path="mem.jsonl"),  # learns across runs
        parallel_tools=True,                   # run independent calls concurrently
    )
    .budget(max_tokens=200_000, max_cost_usd=5.0)
    .resilient(retries=2, timeout=30)
    .checkpoint("run.ckpt")                    # crash-safe + resumable
    .observe()                                 # live dashboard at /observability
    .build()
)

Govern an agent you already have

Antraft is framework-agnostic. Wrap any agent that proposes actions:

existing.py
await Antraft.guard(my_agent, my_tools).allow(["search"]).deny(["shell"]).run()
Next →
Architecture