langchain-verathos 0.1.0

LangChain integration for Verathos — verified LLM inference on Bittensor

langchain-verathos

LangChain integration for Verathos — verified LLM inference on Bittensor.

Every inference response from Verathos is backed by cryptographic proofs (ZK sumcheck + Merkle commitments). Your LangChain chains can verify that the declared model was executed faithfully — no output substitution, no bait-and-switch.

Installation

pip install langchain-verathos

Quick Start

from langchain_verathos import ChatVerathos

# Set your API key (or export VERATHOS_API_KEY=...)
llm = ChatVerathos(api_key="vrt_sk_...")

# "auto" picks the best available model (this is the default)
llm = ChatVerathos(model="auto")

msg = llm.invoke("Explain zero-knowledge proofs in one paragraph.")
print(msg.content)

# Proof verification metadata is on every response
print(msg.response_metadata["proof_verified"])  # True
print(msg.response_metadata["timing"])          # {"inference_ms": ..., "prove_ms": ...}
print(msg.response_metadata["proof_details"])   # {"challenged_layers": [...], ...}

Why not just ChatOpenAI(base_url=...)?

The Verathos API is OpenAI-compatible, so you can use plain ChatOpenAI:

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    base_url="https://api.verathos.ai/v1",
    api_key="vrt_sk_...",
    model="auto",
)
msg = llm.invoke("Hello!")
print(msg.content)  # works fine

This works for basic chat, but you lose the proof metadata. The OpenAI SDK parses responses into typed Pydantic models, and while it preserves extra fields internally, ChatOpenAI doesn't extract them into response_metadata.

ChatVerathos adds:

Feature	ChatOpenAI	ChatVerathos
Chat completions	Yes	Yes
Streaming	Yes	Yes
Tool calling	Yes	Yes
Structured output	Yes	Yes
Proof metadata in response_metadata	No	Yes
Proof metadata in additional_kwargs	No	Yes
VERATHOS_API_KEY env var	No	Yes
model="auto" default	No	Yes
list_models() discovery	No	Yes
LangSmith provider tag verathos	No	Yes

Features

Automatic model selection

Verathos supports model="auto" which automatically selects the best available model from the network. This is the default for ChatVerathos.

llm = ChatVerathos()  # model="auto" by default

Proof verification metadata

Every response includes cryptographic proof metadata:

msg = llm.invoke("What is 2+2?")

# Top-level proof status
msg.response_metadata["proof_verified"]  # True or False

# Detailed timing breakdown
msg.response_metadata["timing"]
# {
#     "inference_ms": 1234.5,
#     "prove_ms": 567.8,
#     "verify_ms": 89.0,
#     ...
# }

# Detailed proof information (when include_proof=True, the default)
msg.response_metadata["proof_details"]
# {
#     "challenged_layers": [3, 17, 28, 41],
#     "total_layers": 48,
#     "beacon_valid": True,
#     "detection_prob_single": 0.0816,
#     "detection_prob_10": 0.5765,
#     "detection_prob_100": 1.0,
#     "proof_size_kb": 12.4,
#     "sampling_active": True,
#     "moe_info": {"is_moe": True, ...},
#     ...
# }

The same metadata is also available in msg.additional_kwargs for convenient programmatic access:

if msg.additional_kwargs.get("proof_verified"):
    print("Response is cryptographically verified!")

Streaming with proof metadata

In streaming mode, proof metadata arrives on the final chunk:

full = None
for chunk in llm.stream("Write a haiku about verification."):
    print(chunk.text, end="", flush=True)
    full = chunk if full is None else full + chunk

print()
# Proof metadata is available after accumulation
print(full.response_metadata.get("proof_verified"))

Model discovery

# List all available models
models = ChatVerathos.list_models()
for m in models:
    print(f"{m['id']:40s}  {m.get('owned_by', '')}")

# Just the model IDs
ids = ChatVerathos.list_model_ids()
# ['Qwen/Qwen3-30B-A3B', 'meta-llama/Llama-3.3-70B-Instruct', ...]

Async support

All async methods work out of the box:

msg = await llm.ainvoke("Hello!")

async for chunk in llm.astream("Hello!"):
    print(chunk.text, end="")

Tool calling and structured output

Since ChatVerathos extends ChatOpenAI, all tool calling and structured output features work:

from pydantic import BaseModel

class Answer(BaseModel):
    reasoning: str
    answer: int

structured_llm = llm.with_structured_output(Answer)
result = structured_llm.invoke("What is 15 * 23?")
print(result.answer)  # 345

Disabling proof details

To save bandwidth, you can disable the detailed proof metadata (you'll still get proof_verified and timing):

llm = ChatVerathos(include_proof=False)

Configuration

Environment variables

Variable	Description
VERATHOS_API_KEY	API key for authentication

Constructor parameters

Parameter	Default	Description
model	"auto"	Model name or "auto" for automatic selection
api_key	VERATHOS_API_KEY env	API key
base_url	https://api.verathos.ai/v1	API base URL
include_proof	True	Include detailed proof metadata
temperature	None	Sampling temperature
max_tokens	None	Maximum tokens to generate
streaming	False	Enable streaming by default

All standard ChatOpenAI parameters (timeout, max_retries, etc.) are also supported.

Getting an API key

1. Visit verathos.ai to create an account
2. Deposit TAO or USDC to get credits
3. Generate an API key from the dashboard

Alternatively, Verathos supports x402 pay-per-request with USDC on Base — no API key or deposit needed.

Architecture

Verathos runs on Bittensor, a decentralized AI network. Miners serve open-weight models with cryptographic proofs. Validators verify proofs and set reputation scores. The Verathos proxy routes your requests to the highest-scoring miners.

Your App -> ChatVerathos -> api.verathos.ai -> Validator Proxy -> Miner (GPU)
                                                    |
                                              Proof Verification
                                              (ZK sumcheck + Merkle)

Every response proves:

• Model integrity: The exact declared model weights were used (Merkle commitment against on-chain root)
• Computation integrity: Matrix multiplications were executed correctly (ZK sumcheck protocol)
• Output binding: The returned text matches the proven computation (SHA-256 commitment)
• Input binding: The prompt you sent is the prompt that was executed (embedding proof)