metalins 0.4.2

Continuous behavioral verification for AI agents in production — detect drift, clones, and prompt injection.

metalins

Continuous behavioral verification for AI agents in production.

Your agents in production are black boxes. Metalins verifies they're still the same agents you deployed — same model, same behavior, continuously. It's the behavioral verification layer in the Zero Trust stack for AI agents.

How it works

1. The SDK hashes your agent's inputs and outputs locally — raw prompts and responses never leave your infrastructure.
2. Signed hashes are sent to api.metalins.ai, where the behavioral engine runs.
3. The engine returns a continuous verification status: verified, caution, or not_verified.

Your data stays in your infra. We only see fingerprints.

Install

pip install metalins

Quick start

Three lines to start verifying your agent:

import metalins

agent = metalins.Agent(api_key="ml_live_...", name="my-agent")
agent.start()

# Log each turn — hashing happens locally, automatically
agent.log(input=user_message, output=agent_reply)

# Check verification status at any time
status = agent.get_status()  # "verified" | "caution" | "not_verified"

Or as a context manager:

with metalins.Agent(api_key="ml_live_...", name="my-agent") as agent:
    agent.log(input=user_message, output=agent_reply)

Get your API key at metalins.ai.

Integrations

LangChain

from metalins import Agent
from metalins.integrations.langchain import MetalinsCallbackHandler

agent = Agent(api_key="ml_live_...", name="my-bot").start()
handler = MetalinsCallbackHandler(agent)

chain.invoke(user_input, config={"callbacks": [handler]})

Every chain and LLM call is logged automatically — no manual agent.log() needed.

FastAPI / Starlette

import metalins
from metalins.integrations.fastapi import MetalinsMiddleware

agent = metalins.Agent(api_key="ml_live_...", name="my-api").start()
app.add_middleware(MetalinsMiddleware, agent=agent)

Every request/response pair is logged automatically. Bodies are hashed locally and never buffered in full (1 MiB cap by default). Skip noisy endpoints with exclude_paths=["/health"].

Anthropic SDK

import metalins

agent = metalins.Agent(api_key="ml_live_...", name="my-claude-agent").start()

with metalins.trace(agent):
    response = client.messages.create(...)

Or use the @metalins.monitor decorator on any function that calls the Anthropic SDK.

What leaves your infrastructure

Only hashed fingerprints — never raw text:

What we receive	What stays with you
SHA-256 hash of input	Raw prompt text
SHA-256 hash of output	Raw response text
Timestamp + agent ID	Your users' data
HMAC-signed event chain	Your model config

The behavioral engine compares fingerprint patterns over time. It does not reconstruct your prompts or responses.

State persistence

The SDK persists the agent session (ID, secret, hash chain) to ~/.metalins/<name>.json with 0600 permissions by default. To store it elsewhere — a database, a secrets manager — pass any object with load() and save():

agent = metalins.Agent(api_key="ml_live_...", name="my-bot", store=my_store)

License

Apache 2.0. See LICENSE.