Govern your AI agent in ~15 minutes

Create your account & get a key

Sign up and grab your API key from the dashboard. Agent governance draws on the same prepaid credits as chat — no new wallet, no new SKU.

Install the SDK

One small dependency. Drop in the two adapter files (or install the package once it's published).

# the binding's only hard deps
pip install httpx langchain-core

# vendor these two files into your project:
#   ark_agent_governance.py   (the over-the-wire client)
#   langchain_ark.py          (the LangChain / LangGraph adapter)

Wrap your tools — start in audit mode

Point a session at the gateway, wrap your existing tools, and attach the recorder. Nothing in your agent's logic changes — ARK just observes and records every step.

from ark_agent_governance import ARKAgentSession
from langchain_ark import ark_guard_tools, ARKCallbackHandler

ark = ARKAgentSession(
    base_url="https://arkgov.io",        # or your self-host URL
    api_key="sk-ark-...",                  # from your dashboard
    goal="book the customer's refund",
    mode="audit",                          # START here: record everything, block NOTHING
)

safe_tools = ark_guard_tools(my_tools, ark)          # gate-before-run + record-after
agent = create_agent(model="...", tools=safe_tools)

Run it — watch, don't block

Run your agent normally. ARK classifies each step by blast-radius (read · write · irreversible), records it to a hash-chained trajectory, and shows what it would have blocked — without stopping anything. This is how you earn trust before enforcing.

result = agent.invoke(task, config={"callbacks": [ARKCallbackHandler(ark)]})
# every step is now graded + recorded; in audit mode nothing is blocked

Prove the run

One call returns the tamper-evident chain-of-custody for the whole trajectory — the record a regulator or your own risk team can re-walk. Any later edit, deletion, or reorder breaks the chain.

print(ark.verify())   # {ok: true, entries: N, head: "..."} — survives restarts

Flip to enforce — when your false-block rate is ~0

Once audit mode shows ARK isn't flagging good actions, switch on enforcement. Now an irreversible, ungrounded action is blocked before it runs, and an action that needs sign-off routes to your human callback. ARK only ever blocks the irreversible-and-ungrounded tier, so a benign or grounded call is never auto-stopped.

ark = ARKAgentSession(
    base_url="https://arkgov.io", api_key="sk-ark-...",
    goal="book the customer's refund",
    mode="enforce",                                       # block / human-approve enforced
    on_approval=lambda tool, args, decision: ask_a_human(tool, args),
)