Python SDK (pip install xybern)

Python SDK, `pip install xybern`¶

The Xybern SDK is the frictionless way to bring an existing agent stack under governance. Install once and it discovers every AI agent in your process, gives each a cryptographic identity, registers them to your workspace, and, when you turn enforcement on, authorises every action before it executes.

pip install xybern
xybern login

from xybern import auto
auto.connect()

✓ Detected frameworks: CrewAI, LangGraph, 2 MCP servers
✓ Found 12 agents · 48 tools
✓ Registered to workspace "Acme Corp" (each issued a cryptographic identity)
✓ Mode: OBSERVE — actions logged, nothing blocked yet
   → run `xybern enforce on` to start authorising

Where to add it¶

Add these two lines once, at your application's startup, before your agents run. That's the whole integration; you don't sprinkle it around your code.

from xybern import auto
auto.connect()
# ... then your normal code: build agents, run the crew/graph, serve requests, etc.

Your setup	Where to put it
Agent script (CrewAI / LangChain / OpenAI Agents / LlamaIndex)	Top of your main file, before you create agents
FastAPI	Right after `app = FastAPI()` (or in startup)
Celery	In the Celery app module (runs in each worker)
MCP server	Top of the server entry, before it serves tools
Django	In an `AppConfig.ready()` or `wsgi.py`
Jupyter notebook	The first cell

# Example — CrewAI
from xybern import auto
auto.connect()                       # ← here, first

from crewai import Agent, Crew
researcher = Agent(role="Researcher", ...)   # auto-registered with an identity

Two things to know

Multi-process (gunicorn, multiple Celery workers): it must run in each worker process, so put it in the module every worker imports (your app factory / worker bootstrap), not a one-off script.
Earlier is better, but it's forgiving: agents created after connect() are registered on creation; ones that already exist are still found by the startup scan and their actions still intercepted.

Auto-discovery¶

The SDK never scans your machine, it inspects what's importable in the running process and hooks the creation + execution points of supported frameworks:

LangChain · CrewAI · OpenAI Agents SDK · MCP servers · LangGraph · AutoGen · Semantic Kernel · LlamaIndex (plus FastAPI / Celery). Every detector is defensive: a version it doesn't recognise degrades to "presence detected" rather than touching your code, so installing the SDK can never break your app.

As your application creates agents and tools, they appear automatically in the Xybern dashboard, each registered with its own identity.

Observe-first, fail-open¶

The default mode is OBSERVE: actions are logged and inventoried, but nothing is ever blocked. Flip to enforcement only when you're ready:

auto.connect(mode="enforce")   # or: xybern enforce on

In enforce mode each action is authorised before execution (allow / block / escalate). If Xybern is unreachable the SDK fails open by default, so it can never take your agents down. Action content is sent as a SHA-256 hash by default, not raw payloads.

Logging in¶

Method	How
Device code	`xybern login`, opens a browser, approve and pick a workspace; a scoped key is minted and saved to `~/.xybern/credentials.json`
API key	`xybern login --api-key xb_...`, or `export XYBERN_API_KEY=xb_...`, or `auto.connect(api_key="xb_...")`

CLI¶

xybern login [--api-key xb_...]
xybern agents          # dry-run discovery — what would be found
xybern doctor          # per-framework enforcement coverage (run in your project env)
xybern status
xybern enforce on|off
xybern logout

Configuration¶

Option	Default	Meaning
`mode`	`observe`	`observe` (log only) or `enforce`
`fail_open`	`true`	allow actions through if Xybern is unreachable (enforce mode)
`redact`	`true`	send content hashes instead of raw content
`frameworks`	all detected	restrict instrumentation to specific frameworks

Every authorised action is recorded in your Provenance Vault, signed, policy-bound, and externally anchored.

Python SDK (pip install xybern)