helix-framework 0.5.4

Production-grade AI agent framework — cost governance, memory, caching, multi-agent teams, and built-in eval

Helix

A Python framework for building production AI agents.

Helix gives you agents that actually behave in production: hard budget limits, semantic caching for repeated queries, opt-in persistent memory (SQLite, zero extra dependency), MCP tool support, agent handoffs, multi-agent teams, YAML-based task pipelines, a LangGraph-compatible StateGraph, and a 6-scorer eval suite. It works out of the box with OpenAI, Anthropic, Gemini, Groq, Mistral, and 8 other providers.

The import helix API is intentionally close to what you already know from AutoGen, CrewAI, and LangGraph, but with the production layer those frameworks leave to you: cost governance, caching, memory, observability, and safety controls.

Table of Contents

• Installation
• Quickstart
• Agents
• Class-Based Agents
• Preset Agents
• Agent Pipelines
• Tools
• MCP Tools
• Tasks and Pipelines
• YAML Configuration
• Multi-Agent Teams
• Handoffs
• Group Chat
• Workflows
• StateGraph
• Sessions
• Budget Enforcement
• Guardrails
• Event Hooks
• Evaluation
• Framework Adapters
• CLI
• Architecture
• Supported Providers
• Contributing

---

Installation

pip install helix-framework                        # core only (pydantic required)
pip install "helix-framework[gemini]"              # + Google Gemini (free tier available)
pip install "helix-framework[openai,anthropic]"    # + OpenAI and Anthropic
pip install "helix-framework[all]"                 # all providers

From source:

git clone https://github.com/sarcasticdhruv/helix-agent
cd helix-agent
pip install -e ".[all]"

API key setup

The easiest way is the persistent config store:

helix config set GOOGLE_API_KEY    "AIza..."    # Gemini, free tier works fine
helix config set OPENAI_API_KEY    "sk-..."
helix config set ANTHROPIC_API_KEY "sk-ant-..."

Keys are saved to ~/.helix/config.json. Helix picks the best available model automatically when multiple keys are set.

Or use environment variables directly:

# Linux / macOS
export GOOGLE_API_KEY="AIza..."

# Windows PowerShell
$env:GOOGLE_API_KEY = "AIza..."

---

Quickstart

import helix

agent = helix.Agent(
    name="Researcher",
    role="Research analyst",
    goal="Find accurate, cited answers.",
)

result = helix.run(agent, "What is quantum entanglement?")
print(result.output)
print(f"Cost:  ${result.cost_usd:.4f}")
print(f"Steps: {result.steps}")

For the fastest possible start, use helix.quick() — no config objects needed:

import helix

agent = helix.quick("You are a concise Python tutor.", budget_usd=0.10)
result = helix.run(agent, "Explain list comprehensions.")
print(result.output)

Inside an async function, call run_async or agent.run directly:

import asyncio
import helix

async def main():
    agent = helix.Agent(
        name="Researcher",
        role="Research analyst",
        goal="Find accurate answers.",
    )
    result = await agent.run("What is quantum entanglement?")
    print(result.output)

asyncio.run(main())

helix.quick() parameters:

Parameter	Description
system_prompt	The agent's purpose as plain instructions
name	Agent name shown in traces (default "Agent")
model	Model string, e.g. "gpt-4o". Auto-detected if omitted
tools	List of @helix.tool-decorated functions
budget_usd	Hard spend cap per run (default 0.10)
on_event	Optional async/sync event callback (see Event Hooks)

---

Agents

import helix

agent = helix.Agent(
    name="Analyst",
    role="Senior data analyst",
    goal="Analyze datasets and produce concise summaries.",

    # Optional: rich background context that shapes agent behaviour
    backstory=(
        "You have 8 years of experience in financial data analysis. "
        "You prefer bullet-point summaries over long prose."
    ),

    # Model selection with automatic fallback
    model=helix.ModelConfig(
        primary="gpt-4o",
        fallback_chain=["gpt-4o-mini", "gemini-2.0-flash"],
        temperature=0.3,
    ),

    # Hard cost limit
    budget=helix.BudgetConfig(budget_usd=1.00),
    mode=helix.AgentMode.PRODUCTION,

    # Memory
    memory=helix.MemoryConfig(short_term_limit=20),

    # Semantic caching (cost reduction on repeated/similar queries)
    cache=helix.CacheConfig(enabled=True, semantic_threshold=0.92),
)

result = helix.run(agent, "Summarize last quarter's sales trends.")

AgentResult fields: output, cost_usd, steps, model_used, cache_hits, cache_savings_usd, tool_calls, run_id, duration_s, trace.

Agents also expose LangChain-compatible aliases: agent.invoke(task) (sync) and await agent.ainvoke(task) (async), both equivalent to helix.run() / await agent.run().

---

Class-Based Agents

The @helix.agent decorator turns any class into an Agent factory. Tools become methods decorated with @helix.tool, and the class docstring becomes the system prompt.

import helix

@helix.agent(model="claude-sonnet-4-6", budget_usd=2.00)
class WebResearcher:
    """
    You are an expert web researcher.
    Find accurate, up-to-date information and always cite sources.
    """

    @helix.tool(description="Search the web for recent information.")
    async def search(self, query: str) -> list[dict]:
        from helix.tools.builtin import web_search
        return await web_search(query)

    @helix.tool(description="Fetch and read a URL.")
    async def fetch(self, url: str) -> str:
        from helix.tools.builtin import fetch_url
        result = await fetch_url(url)
        return result.get("content", "")

# The decorator returns a factory; call it to get an Agent instance
researcher = WebResearcher()
result = helix.run(researcher, "Latest AI safety research 2026")

@helix.agent options:

Parameter	Description
model	LLM model string. Auto-detected if omitted
budget_usd	Spending cap per run (default 0.50)
mode	"explore" (default) or "production"
name	Override agent name; defaults to class name
backstory	Rich background context injected into system prompt

---

Preset Agents

helix.presets provides nine ready-made agent factories so you can start in one line:

from helix.presets import web_researcher, writer, coder, summariser

# Single agent
result = helix.run(web_researcher(), "Top AI papers this week")

# Code generation
result = helix.run(coder("TypeScript"), "Write a UUID v4 generator")

# Pipe agents together with |
research_and_write = web_researcher() | writer()
result = research_and_write.run_sync("Write a report on quantum computing")

Available presets:

Factory	Description
web_researcher(budget_usd, model, max_results)	Web search + URL fetching
writer(style, budget_usd)	Polished prose from bullet points or research
summariser(style, budget_usd)	Compress long text into a summary
fact_checker(budget_usd)	Verify claims against web sources
coder(language, budget_usd, allow_file_io)	Code generation and debugging
code_reviewer(language, budget_usd)	Bug, style, and security review
data_analyst(budget_usd)	Statistical analysis with calculator tool
api_agent(base_url, auth_token, budget_usd)	REST API orchestration
assistant(domain, budget_usd)	General-purpose assistant for any domain

---

Agent Pipelines

Use the | operator to wire agents into a sequential pipeline. Each agent's output becomes the next agent's input.

from helix.presets import web_researcher, summariser, writer

# Build with | operator
pipeline = web_researcher() | summariser() | writer(style="blog post")
result = pipeline.run_sync("Quantum computing advances in 2026")
print(result.output)
print(f"Cost: ${result.cost_usd:.4f}")

Or use helix.chain() for a more explicit form:

pipeline = helix.chain(web_researcher(), summariser(), writer())
result = pipeline.run_sync("Quantum computing advances in 2026")

AgentPipeline also exposes:

• await pipeline.run(task) — async version
• pipeline.agents — list of Agent instances in the chain

---

Tools

import helix

@helix.tool(
    description="Search the web for current information.",
    timeout=15.0,
    retries=2,
)
async def web_search(query: str, max_results: int = 5) -> list:
    # your implementation here
    return [{"title": "...", "url": "...", "snippet": "..."}]


@helix.tool(description="Read a file from disk.")
async def read_file(path: str) -> str:
    with open(path) as f:
        return f.read()


agent = helix.Agent(
    name="Researcher",
    role="Research analyst",
    goal="Find answers using web search.",
    tools=[web_search, read_file],
)

result = helix.run(agent, "What are the latest AI headlines?")

Built-in tools (13 included):

import helix.tools.builtin  # registers tools globally

# web_search, fetch_url, read_file, write_file, list_directory,
# calculator, json_query, get_datetime, get_env,
# text_stats, extract_urls, sleep, execute_python

The execute_python tool runs sandboxed Python code in an isolated subprocess. Dangerous modules (subprocess, os.system, ctypes, pty, multiprocessing) are blocked. Returns {"success", "stdout", "stderr", "returncode"} with a configurable timeout (default 15 s).

from helix.tools.builtin import execute_python

# Use inside an agent that needs to run arbitrary Python
agent = helix.Agent(
    name="Calculator",
    role="Python executor",
    goal="Run Python snippets and return results.",
    tools=[execute_python],
)

Use helix.discover_tools() to list every tool registered in the global registry (built-ins + any @helix.tool functions loaded at import time).

Breaking change (v0.5): agents no longer automatically get every globally-registered tool. Pass the tools you want explicitly via tools=[...], or opt in to the old behavior with Agent(..., inherit_global_tools=True).

---

MCP Tools

Connect to any Model Context Protocol server and use its tools like any other Helix tool. Requires pip install "helix-framework[mcp]".

import helix
from helix.tools.mcp import MCPToolSource

async def main():
    async with MCPToolSource(command="npx", args=["-y", "@some/mcp-server"]) as tools:
        agent = helix.Agent(
            name="Bot",
            role="Assistant",
            goal="Use the connected MCP tools to help the user.",
            tools=tools,
        )
        result = await agent.run("...")
        print(result.output)

Without the context manager (when the connection needs to outlive one call):

source = MCPToolSource(command="python", args=["-m", "my_mcp_server"])
tools = await source.connect()
# ... use tools across multiple agent runs ...
await source.close()

---

Tasks and Pipelines

Tasks are first-class declarative units of work. They chain outputs together, support output validation with guardrails, and can write results to files. This is the Helix equivalent of CrewAI's Task + crew.kickoff().

import helix

researcher = helix.Agent(
    name="Researcher",
    role="Research analyst",
    goal="Find accurate information on {topic}.",
    backstory="You specialize in academic and technical research.",
)
writer = helix.Agent(
    name="Writer",
    role="Technical writer",
    goal="Write clear articles on {topic}.",
)

research = helix.Task(
    description="Research the latest advances in {topic}.",
    expected_output="A list of 5 key findings with sources.",
    agent=researcher,
)
article = helix.Task(
    description="Write a 3-paragraph article based on the research.",
    expected_output="A well-structured article, no jargon.",
    agent=writer,
    context=[research],        # automatically receives research output
    output_file="article.md",  # saved to disk when done
)

pipeline = helix.Pipeline(tasks=[research, article])
result = pipeline.kickoff(inputs={"topic": "quantum computing"})
print(result.final_output)
print(f"Total cost: ${result.total_cost_usd:.4f}")

Task options:

Parameter	Description
context	List of Tasks whose outputs are passed as context
output_schema	Pydantic model for structured output
guardrail	Validation function or string description
guardrails	List of validation functions (chained)
guardrail_max_retries	How many times to retry on validation failure (default 3)
output_file	Path to write the task output
async_execution	Run this task concurrently with others
callback	Called with TaskOutput after completion
markdown	Instruct the agent to format output as Markdown

Validation with guardrails:

from helix import Task, TaskOutput

def must_be_under_300_words(result: TaskOutput):
    words = len(result.raw.split())
    if words > 300:
        return False, f"Too long: {words} words (max 300)"
    return True, result.raw

task = helix.Task(
    description="Write a short summary of {topic}.",
    expected_output="A summary under 300 words.",
    agent=writer,
    guardrail=must_be_under_300_words,
    guardrail_max_retries=2,
)

You can also pass a plain string and Helix uses the agent's own LLM to validate:

task = helix.Task(
    description="Write a product description for {product}.",
    expected_output="A concise, professional product description.",
    agent=writer,
    guardrail="Must be professional, under 100 words, and avoid superlatives.",
)

Accessing task output:

result = pipeline.kickoff(inputs={"topic": "AI safety"})

for task_output in result.task_outputs:
    print(f"Task:  {task_output.summary}")
    print(f"Raw:   {task_output.raw}")
    if task_output.pydantic:
        print(f"Model: {task_output.pydantic}")

---

YAML Configuration

Define agents and tasks in YAML files for cleaner project structure:

# agents.yaml
researcher:
  role: Senior Research Analyst
  goal: Find cutting-edge developments in {topic}.
  backstory: You work at a leading tech think tank with access to academic databases.

writer:
  role: Content Strategist
  goal: Write engaging, accurate articles about {topic}.
  backstory: You have 5 years of experience writing technical content for developers.

# tasks.yaml
research_task:
  description: Research the latest developments in {topic}.
  expected_output: A structured report with at least 5 key findings.
  agent: researcher

write_task:
  description: Write a concise article based on the research.
  expected_output: A 3-paragraph article written for a developer audience.
  agent: writer
  context: [research_task]
  output_file: output/article.md

import helix

pipeline = helix.from_yaml(
    "agents.yaml",
    "tasks.yaml",
    inputs={"topic": "large language models"},
)
result = pipeline.kickoff()
print(result.final_output)

Or use the lower-level helpers:

from helix.core.yaml_config import load_agents, load_tasks, load_pipeline

agents   = load_agents("agents.yaml", inputs={"topic": "LLMs"})
tasks    = load_tasks("tasks.yaml", agents, inputs={"topic": "LLMs"})
pipeline = load_pipeline(tasks)
result   = pipeline.kickoff()

---

Multi-Agent Teams

Teams coordinate multiple agents with three execution strategies.

import helix

searcher = helix.Agent(name="Searcher", role="Web researcher",   goal="Find sources.")
analyst  = helix.Agent(name="Analyst",  role="Data analyst",     goal="Analyze data.")
writer   = helix.Agent(name="Writer",   role="Technical writer", goal="Write reports.")

# sequential: searcher output feeds into analyst, then into writer
team = helix.Team(
    name="research-team",
    agents=[searcher, analyst, writer],
    strategy="sequential",
    budget_usd=5.00,
)

result = team.run_sync("Write a report on renewable energy trends.")
print(result.final_output)
print(f"Total cost: ${result.total_cost_usd:.4f}")

Strategies:

• sequential - each agent receives the previous agent's output as its input
• parallel - all agents run on the same input concurrently, outputs returned as a list
• hierarchical - a lead agent decomposes the task and delegates subtasks to specialists

lead = helix.Agent(name="Lead", role="Project lead", goal="Decompose and delegate tasks.")

team = helix.Team(
    name="product-team",
    agents=[searcher, analyst, writer],
    strategy="hierarchical",
    lead=lead,
)

---

Handoffs

Give an agent a list of handoffs and it can transfer the conversation to a specialist mid-run — visible to the model as a normal tool call (transfer_to_<name>), not a hidden orchestration decision.

import helix

billing = helix.Agent(
    name="BillingAgent",
    role="Billing specialist",
    goal="Handle billing questions, invoices, refunds, and payment issues.",
)
triage = helix.Agent(
    name="Triage",
    role="Customer support triage agent",
    goal="Figure out what the customer needs. Transfer billing questions to the billing agent.",
    handoffs=[billing],
)

result = await triage.run("My invoice #4471 shows double charges, can you help?")
print(result.output)          # produced by BillingAgent, not Triage
print(result.handoff_chain)   # ["Triage"] — the path this run took before reaching its answer
print(result.cost_usd)        # includes cost from every agent in the chain

handoffs can be combined with guardrails and any other Agent option. Chained handoffs (A → B → C) accumulate in handoff_chain in order.

---

Group Chat

Group chat puts multiple agents in a shared multi-turn conversation. This is Helix's equivalent of AutoGen's GroupChat.

import asyncio
import helix

ceo    = helix.ConversableAgent(name="CEO",    role="CEO",    goal="Make strategic decisions.")
cto    = helix.ConversableAgent(name="CTO",    role="CTO",    goal="Assess technical risk.")
lawyer = helix.ConversableAgent(name="Lawyer", role="Lawyer", goal="Flag compliance issues.")

chat = helix.GroupChat(
    agents=[ceo, cto, lawyer],
    max_rounds=6,
    speaker_selection="round_robin",  # or "auto", "random", or a callable
    termination_keyword="AGREED",
)

async def main():
    result = await chat.run("Should we migrate our core product to microservices?")
    print(result.transcript())
    print(f"Rounds: {result.rounds}, Cost: ${result.total_cost_usd:.4f}")

asyncio.run(main())

Speaker selection:

Value	Behavior
round_robin	Agents speak in order (default)
auto	A coordinator LLM picks the most relevant next speaker
random	Random selection each round
callable	fn(agents, history) -> Agent

Termination:

chat = helix.GroupChat(
    agents=[...],
    max_rounds=10,
    termination_keyword="FINAL ANSWER",
    termination_fn=lambda msgs: len(msgs) > 8,
)

Human in the loop:

human = helix.HumanAgent(name="You")   # prompts the terminal each turn

chat = helix.GroupChat(
    agents=[agent1, agent2, human],
    max_rounds=5,
)

---

StateGraph

helix.StateGraph is a LangGraph-compatible directed graph engine for building complex agentic pipelines with cycles, conditional branching, and checkpoint persistence.

import helix
from typing import TypedDict

class State(TypedDict):
    topic: str
    draft: str
    ready: bool

researcher = helix.presets.web_researcher()
writer     = helix.presets.writer()

async def research_node(state: State) -> dict:
    result = await researcher.run(state["topic"])
    return {"draft": result.output}

async def write_node(state: State) -> dict:
    result = await writer.run(state["draft"])
    return {"draft": result.output, "ready": True}

def router(state: State) -> str:
    return helix.END if state.get("ready") else "write"

graph = (
    helix.StateGraph(State)
    .add_node("research", research_node)
    .add_node("write", write_node)
    .add_edge("research", "write")
    .add_conditional_edges("write", router, {"write": "write", helix.END: helix.END})
    .set_entry_point("research")
    .compile()
)

result = graph.run_sync({"topic": "Quantum computing in 2026", "draft": "", "ready": False})
print(result["draft"])

StateGraph API:

Method	Description
.add_node(name, fn)	Register an async or sync callable as a graph node
.add_edge(a, b)	Unconditional edge from node a to node b
.add_conditional_edges(node, fn, mapping)	Routing function determines next node
.set_entry_point(node)	First node executed
.set_finish_point(node)	Node that signals graph completion
.compile(checkpoint_dir=...)	Returns CompiledGraph

CompiledGraph execution:

result = graph.run_sync(initial_state)       # synchronous
result = await graph.run(initial_state)      # async
result = await graph.ainvoke(initial_state)  # LangChain-compatible alias
async for state in graph.stream(initial_state):  # step-by-step streaming
    print(state)

helix.END and helix.START are exported directly from the top-level helix namespace.

Pass checkpoint_dir=".helix/checkpoints" to .compile() to save state after every node, enabling resume-after-crash for long-running graphs.

---

Workflows

Workflows are step-based directed pipelines with retry, timeout, fallback, and branching.

import helix

@helix.step(name="search", retry=2, timeout_s=10.0)
async def search_step(query: str) -> list:
    return []  # your search implementation

@helix.step(name="summarise")
async def summarise_step(results: list) -> str:
    return "\n".join(str(r) for r in results)

pipeline = (
    helix.Workflow("research-pipeline")
    .then(search_step)
    .then(summarise_step)
    .with_budget(2.00)
)

result = pipeline.run_sync("quantum computing trends 2025")
print(result.final_output)

---

Sessions

Sessions give an agent persistent memory across multiple turns.

import asyncio
import helix

async def main():
    agent = helix.Agent(name="Bot", role="Assistant", goal="Help users.")
    session = helix.Session(agent=agent)
    await session.start()

    r1 = await session.send("My name is Alice.")
    r2 = await session.send("What is my name?")   # remembers: Alice
    print(r2.output)

    await session.end()

asyncio.run(main())

---

Event Hooks

Attach an on_event callback to any agent to receive live telemetry without any extra config — no trace files, no external services.

import helix
from helix.core.hooks import HookEvent

async def my_hook(event: HookEvent) -> None:
    if event.type == "tool_call":
        print(f"  → {event.data['tool_name']}({event.data['args']})")
    elif event.type == "step_end":
        print(f"  ✓ step {event.step} — ${event.cost_so_far:.4f} spent")
    elif event.type == "llm_call":
        print(f"  [LLM] {event.data['model']}")
    elif event.type == "cache_hit":
        print(f"  [CACHE] saved ${event.data['saved_usd']:.4f}")

agent = helix.Agent(
    name="Researcher",
    role="Research analyst",
    goal="Find information.",
    on_event=my_hook,  # sync or async
)

Event types:

Event	Payload keys
step_start	step
step_end	step, output_preview
llm_call	model, messages
llm_response	model, tokens, finish_reason
tool_call	tool_name, args
tool_result	tool_name, result_preview
tool_error	tool_name, error
cache_hit	similarity, saved_usd
done	output_preview, steps, cost_usd
error	error

Hook errors are silently swallowed so they never affect agent execution. Both sync and async callables are supported.

---

Budget Enforcement

import helix

agent = helix.Agent(
    name="Bot",
    role="Assistant",
    goal="Help users.",
    budget=helix.BudgetConfig(
        budget_usd=0.50,
        warn_at_pct=0.8,
        strategy=helix.BudgetStrategy.DEGRADE,  # step down to cheaper model instead of stopping
    ),
    mode=helix.AgentMode.PRODUCTION,
)

try:
    result = helix.run(agent, "Write a 10,000 word essay on climate change...")
except helix.BudgetExceededError as e:
    print(f"Budget hit: ${e.spent_usd:.4f} of ${e.budget_usd:.4f}")

With BudgetStrategy.DEGRADE, Helix steps down through the fallback chain as the budget depletes rather than stopping outright.

---

Guardrails

Guardrails run on both the incoming task and the model's output — pass built-in guardrail names to Agent(...) directly:

import helix

agent = helix.Agent(
    name="Bot",
    role="Assistant",
    goal="Help users.",
    guardrails=["prompt_injection", "pii_redactor", "length_guard"],
)

Name	Behavior
prompt_injection	Blocks common jailbreak/injection phrasings (instruction override, persona hijack, restriction bypass, "developer mode", system-prompt exfiltration). Heuristic pattern matching, not a trained classifier.
pii_redactor	Redacts email, phone, SSN, credit card, and IP address patterns. Never blocks — cleans and passes through.
length_guard	Blocks responses shorter than min_chars or longer than max_chars (defaults: 1 / 100,000).
keyword_block	Blocks content containing configured keywords (default: none — must be built directly for a custom list).
schema_guard	Validates that JSON-shaped output actually parses as JSON.

A guardrail violation raises helix.errors.GuardrailViolationError.

guardrails=[...] on Agent(...) only supports built-in names with default parameters (build_guardrail_chain under the hood). For a custom-configured guardrail — a specific keyword list, flag instead of block mode — build and call a GuardrailChain directly wherever you need it:

from helix.safety.guardrails import GuardrailChain, KeywordBlockGuard, PromptInjectionGuard

chain = GuardrailChain([
    PromptInjectionGuard(on_fail="flag"),  # let it through but record the reason
    KeywordBlockGuard(blocked_keywords=["competitor_name"]),
])
result = await chain.check(some_text, context=None)

---

Evaluation

import asyncio
import helix
from helix.eval.suite import EvalSuite
from helix.config import EvalCase

suite = EvalSuite("qa-suite")
suite.add_cases([
    EvalCase(
        name="capital_cities",
        input="What is the capital of France?",
        expected_facts=["Paris"],
        max_cost_usd=0.05,
    ),
    EvalCase(
        name="math",
        input="What is 15% of 240?",
        expected_facts=["36"],
        max_cost_usd=0.05,
    ),
])

async def main():
    agent = helix.Agent(name="Bot", role="Assistant", goal="Answer questions accurately.")
    results = await suite.run(agent, verbose=True)
    print(f"Pass rate:  {results.pass_rate:.0%}")
    print(f"Total cost: ${results.total_cost_usd:.4f}")
    suite.assert_pass_rate(0.90)   # raises AssertionError if below 90%

asyncio.run(main())

The eval suite runs 6 scorers per case: factual accuracy, tool selection, trajectory adherence, cost efficiency, step efficiency, and output quality.

@suite.case decorator:

from helix.eval.suite import EvalSuite
from helix.config import EvalCase

suite = EvalSuite("my-suite")

@suite.case
def capitals():
    return EvalCase(
        input="What is the capital of Germany?",
        expected_facts=["Berlin"],
        max_cost_usd=0.05,
    )

@suite.case
def arithmetic():
    return EvalCase(
        input="What is 25% of 400?",
        expected_facts=["100"],
    )

# suite now has both cases registered; the function name becomes the case name

EvalCase options:

Parameter	Description
input	Task string sent to the agent
expected_facts	Strings that must appear in the output
expected_tools	Tool names the agent is expected to call
expected_trajectory	ExpectedTrajectory for sequence/step constraints
max_steps	Maximum reasoning steps (default 10)
max_cost_usd	Cost cap per case (default 1.00)
pass_threshold	Minimum overall score to pass (default 0.70)
tags	Labels for filtering subsets

---

Framework Adapters

Wrap existing LangChain, CrewAI, or AutoGen code with Helix cost governance:

from langchain_openai import ChatOpenAI
import helix

llm = helix.wrap_llm(ChatOpenAI(model="gpt-4o"), budget_usd=2.00)
# adds budget gate, cost tracking, tracing, and audit log to any LangChain LLM

from langchain.chains import LLMChain
import helix

vchain = helix.from_langchain(LLMChain(...), budget_usd=3.00)
result = await vchain.run(inputs={"input": "Summarise this"})
print(f"Cost: ${vchain.cost_usd:.4f}")

from crewai import Crew
import helix

crew = Crew(agents=[...], tasks=[...])
wrapped = helix.from_crewai(crew, budget_usd=5.00)
result = await wrapped.run(inputs={"topic": "AI trends"})
print(f"Cost: ${wrapped.cost_usd:.4f}")

from autogen import AssistantAgent
import helix

ag_agent = AssistantAgent("assistant", llm_config={...})
wrapped = helix.from_autogen(ag_agent, budget_usd=2.00)
result = await wrapped.run(inputs={"message": "Explain transformers"})

---

CLI

helix doctor                          # check environment and provider keys
helix models                          # list available models with pricing
helix cost --all                      # cost report across all runs
helix trace <run-id>                  # view a run trace
helix trace <run-id> --diff <run-id>  # compare two runs for divergence
helix replay <run-id>                 # interactive failure replay
helix config set KEY value            # set a provider API key

---

Architecture

helix/
├── core/            Agent, ConversableAgent, GroupChat, Task, Pipeline,
│                    Workflow, Team, Session, Tool, StateGraph, AgentPipeline,
│                    Handoff (transfer_to_* tools between agents)
├── presets/         9 ready-made agent factories (web_researcher, coder, writer, …)
├── tools/           13 builtins + MCP client (connect to any MCP server's tools)
├── memory/          Short-term buffer, WAL-backed long-term store, episodic recall,
│                    backends: inmemory (default) | sqlite (persists) | qdrant/pinecone/chroma (not yet implemented)
├── cache/           Semantic cache (tier 1), plan cache (tier 2), prefix cache (tier 3)
├── models/          Router, complexity estimator, 12 provider backends
├── safety/          Cost governor, permission model, guardrails (incl. prompt_injection), HITL, audit log
├── context_engine/  Multi-factor token decay, context compactor, preflight estimator
├── eval/            EvalSuite, 6 scorers, @suite.case decorator, trajectory eval,
│                    regression gate, monitor
├── observability/   Tracer, ghost debug resolver, failure replay
├── adapters/        LangChain, CrewAI, AutoGen + universal LLM wrapper
├── runtime/         Event loop, worker pool, health checks
└── cli/             doctor, models, cost, trace, replay, config, ...

---

Supported Providers

Environment variable	Provider	Notable models	Free tier
GOOGLE_API_KEY	Google Gemini	Gemini 2.5 Flash/Pro, 2.0 Flash	Yes
OPENAI_API_KEY	OpenAI	GPT-4o, GPT-4o-mini, o1, o3, o3-mini	No
ANTHROPIC_API_KEY	Anthropic	Claude Opus 4.6, Sonnet 4.6, Haiku 4.5	No
GROQ_API_KEY	Groq	Llama 3.3-70B, Llama 3.1-8B, Mixtral, Gemma 2	Yes
MISTRAL_API_KEY	Mistral AI	Mistral Large/Small, Codestral, Pixtral	Partial
COHERE_API_KEY	Cohere	Command R+	Partial
TOGETHER_API_KEY	Together AI	200+ open-source models	No
OPENROUTER_API_KEY	OpenRouter	100+ models	Partial
DEEPSEEK_API_KEY	DeepSeek	DeepSeek V3, R1	No
XAI_API_KEY	xAI	Grok	No
PERPLEXITY_API_KEY	Perplexity	Online search models	No
FIREWORKS_API_KEY	Fireworks	Fast open-source inference	No

Set multiple keys and Helix automatically falls back to the next available provider on failure. Ollama (ollama/* or local/*) and any OpenAI-compatible endpoint (Azure, custom base URL) are also supported without environment variable requirements.

---

Framework Comparison

Feature	Helix	AutoGen	CrewAI	LangGraph
Cost Governance	✓ Hard budget limits + semantic caching	✗	✗	✗
Memory	✓ Multi-tier (short-term, episodic, WAL-backed)	Partial	Partial	✗
Semantic Caching	✓ Tier 1 embedding-based cache	✗	✗	✗
Multi-Agent Teams	✓ Handoffs + group chat	✓	✓	✗
YAML Pipelines	✓ Task + workflow YAML	✗	✓	✗
Built-in Eval Suite	✓ 6 scorers + regression gate	✓	✗	✗
MCP Tools	✓ Native MCP client	✗	✗	✗
Guardrails	✓ Prompt injection, HITL, audit log	Partial	Partial	✗
LangGraph Compat	✓ StateGraph, StateGraph.START	✗	✗	Native
Framework Wrappers	✓ LangChain, CrewAI, AutoGen	N/A	✓	✗

When to choose Helix:

• You need hard cost control over multi-step agents (budget overruns are expensive in production)
• Caching matters — repeating queries with semantic similarity wastes money
• You want production observability (traces, eval gates, failure replay)
• Your team already uses CrewAI/AutoGen/LangChain and needs governance overlay

---

FAQ

Is Helix free? Yes. Apache 2.0 license, no paid tier, no usage limits imposed by Helix itself (only your own budget config and provider costs apply).

Does Helix support MCP servers? Yes — Helix ships a native MCP client. Connect to any MCP server and its tools become available to your agents automatically. See MCP Tools.

Can I use Helix with LangChain, CrewAI, or AutoGen agents I already built? Yes. helix.wrap_llm(), helix.from_langchain(), helix.from_crewai(), and helix.from_autogen() wrap existing code from those frameworks with Helix's cost governance, caching, and observability — no rewrite required. See Framework Adapters.

Does Helix work with local models? Yes, via Ollama (ollama/* or local/* model strings) or any OpenAI-compatible endpoint (Azure, custom base URLs), with no environment variable requirements.

How is Helix different from LangGraph? Helix ships a LangGraph-compatible StateGraph (same START/END API) plus things LangGraph doesn't provide out of the box: hard budget limits, semantic caching, multi-tier memory, and a built-in eval suite. See the Framework Comparison table above.

What does "hard budget limits" actually mean? Every agent run can be capped with BudgetConfig(budget_usd=...). If a run would exceed that cap, Helix raises BudgetExceededError instead of continuing to spend — no silent overage.

Does Helix require an API key to try it? No — Google Gemini and Groq both have usable free tiers, and helix doctor will tell you which providers are configured.

---

Contributing

Read CONTRIBUTING.md before opening a PR.

git clone https://github.com/sarcasticdhruv/helix-agent
cd helix-agent
pip install -e ".[dev,gemini]"
pytest tests/

---

Contributors

Name	Role
Dhruv Choudhary	Author and maintainer

---

License

Apache License 2.0. Copyright 2026 Dhruv Choudhary.

See CHANGELOG.md for release history.