trustloop

Python SDK for TrustLoop — governance, audit trail, and kill-switch for AI agents.

Intercept every tool call your agent makes. Log it. Block dangerous actions. Require human approval. Works with LangChain, CrewAI, AutoGen, or any custom Python agent.

Install

pip install trustloop

With async support:

pip install trustloop[async]

With LangChain integration:

pip install trustloop[langchain]

With CrewAI integration:

pip install trustloop[crewai]

Everything:

pip install trustloop[all]

Quick start

Get your free API key at trustloop.live/signup.

from trustloop import TrustLoop

tl = TrustLoop(api_key="tl_your_key_here", agent_name="my-agent")

# Check before running any tool
result = tl.intercept("send_email", {"to": "ceo@bank.com", "body": "..."})
if not result["allowed"]:
    raise RuntimeError(result["message"])

# ... run the tool

Or set your key as an env var and let the SDK find it:

export TRUSTLOOP_API_KEY="tl_your_key_here"
export TRUSTLOOP_AGENT_NAME="my-agent"

tl = TrustLoop()  # reads from env

Usage

Manual intercept

result = tl.intercept("delete_database", {"table": "users"})

# result = {
#   "allowed": False,
#   "status": "BLOCKED",
#   "message": "Matched rule: block destructive database operations"
# }

Auto-raise on block

from trustloop import TrustLoop, TrustLoopBlockedError, TrustLoopPendingError

try:
    tl.intercept("transfer_funds", {"amount": 50000}, raise_if_blocked=True)
except TrustLoopBlockedError as e:
    print(f"Blocked: {e}")
except TrustLoopPendingError as e:
    print(f"Waiting for approval: {e.approval_id}")

@tl.guard() decorator

@tl.guard("send_email")
def send_email(to: str, subject: str, body: str):
    # Only runs if TrustLoop allows it
    ...

@tl.guard()  # uses function name as tool name
def delete_user(user_id: str):
    ...

Async

from trustloop import AsyncTrustLoop

async with AsyncTrustLoop(api_key="tl_...") as tl:
    await tl.intercept("post_tweet", {"text": "Hello world"}, raise_if_blocked=True)

    @tl.guard("send_email")
    async def send_email(to, subject, body):
        ...

LangChain

from trustloop import TrustLoop
from trustloop.integrations.langchain import wrap_tools

tl = TrustLoop(api_key="tl_...", agent_name="langchain-agent")

# Wrap all tools — one line, zero boilerplate
tools = wrap_tools([search_tool, email_tool, db_tool], tl)

agent = create_openai_tools_agent(llm, tools, prompt)

CrewAI

from trustloop import TrustLoop
from trustloop.integrations.crewai import governed_tool

tl = TrustLoop(api_key="tl_...", agent_name="crew-agent")

@governed_tool(tl)
class SendEmailTool(BaseTool):
    name = "send_email"
    description = "Send an email"

    def _run(self, to: str, subject: str, body: str) -> str:
        ...  # only runs if TrustLoop allows it

Governance rules

# Create a rule in plain English
tl.create_rule(
    "Any wire transfer over £10,000 requires human approval",
    action="approve",
    approver_email="cfo@mycompany.com",
)

# Block a tool instantly (kill-switch)
tl.block_tool("drop_table", reason="Emergency: DB ops disabled")

# Unblock
tl.unblock_tool("drop_table")

Audit log

# Get recent calls
logs = tl.get_logs(limit=100, status="BLOCKED")

# Export as CSV
csv = tl.export_logs()

# Stats
stats = tl.get_stats()
print(stats["total"], stats["blocked"])

Human approvals

# List pending
pending = tl.get_pending_approvals()

# Approve or deny programmatically
tl.decide(pending[0]["id"], "approved")

Context manager

with TrustLoop(api_key="tl_...") as tl:
    tl.intercept("my_tool", {...})
# connection closed automatically

MCP (Claude Desktop)

url = TrustLoop.mcp_url("tl_your_key")
# → "https://trustloop-production.up.railway.app/sse?api_key=tl_..."

Paste this into your claude_desktop_config.json:

{
  "mcpServers": {
    "trustloop": { "url": "<paste url here>" }
  }
}

Environment variables

Links

• Dashboard
• Docs
• Sign up free
• npm SDK

License

MIT