Give agents tools.
Define their boundaries.

Claude Code in 2 minutes.

Install PyActuator 0.0.17, the current PyPI release, connect Claude Code over MCP, and refuse one order before the adapter runs. The refusal is written to a local audit file. PyActuator is its own package. This path does not start another Optophi package.

Install the current release.

These commands match PyPI pyactuator 0.0.17, the current release.

python -m pip install 'pyactuator[mcp]==0.0.17'

Register the MCP server with Claude Code.

This is the command in the 0.0.17 guide docs/guides/mcp-server.md. The profile communications is a built-in profile. pyactuator mcp serve --help accepts --profile, and pyactuator tools profiles lists communications.

claude mcp add pyactuator -- pyactuator mcp serve --profile communications

The command adds the PyActuator MCP server and offers the built-in communications profile.

A risky order stops before the adapter runs.

Save the program below as refuse.py, or download the same file, then run it. A limit of 10000 refuses a DEMO order of 100 at 250. The policy gate returns POLICY_DENIED and does not call the broker adapter.

Download refuse.py

python refuse.py

# PyActuator 0.0.17. python -m pip install 'pyactuator[mcp]==0.0.17'
import json
from decimal import Decimal

from pyactuator.core.audit import AuditQueryFilters
from pyactuator.core.policy import LimitNotional
from pyactuator.store.sqlite_audit import SqliteAuditWriter
from pyactuator.sync import invoke, run

audit = SqliteAuditWriter("audit.sqlite")
result = invoke(
    "pyactuator.trading.place_order",
    {
        "symbol": "DEMO",
        "side": "buy",
        "quantity": "100",
        "order_type": "limit",
        "limit_price": "250",
        "client_order_id": "demo-order-001",
    },
    policy_enforcer=LimitNotional(Decimal("10000")),
    audit_writer=audit,
)
record = run(audit.query(AuditQueryFilters(tag="policy_denied", limit=1))).records[0]
print(
    json.dumps(
        {
            "success": result["success"],
            "code": result["error"]["code"],
            "message": result["error"]["message"],
            "adapter_class": result["adapter"]["class"],
            "adapter_real": result["adapter"]["real"],
            "audit_operation": record.operation,
            "audit_success": record.success,
            "audit_tags": list(record.tags),
        },
        indent=2,
    )
)

{
  "success": false,
  "code": "POLICY_DENIED",
  "message": "notional 25000 exceeds limit 10000 (qty=100, price=250)",
  "adapter_class": "MockExecutionClient",
  "adapter_real": false,
  "audit_operation": "pyactuator.trading.place_order",
  "audit_success": false,
  "audit_tags": [
    "policy",
    "policy_denied"
  ]
}

adapter_real is false. The envelope names MockExecutionClient so a mock configuration stays visible. The same run prints two lines on stderr: the analytics extra is not installed, and the policy denial. The order call does not use the analytics extra.

The refusal is a row in the audit file.

refuse.py writes audit.sqlite through SqliteAuditWriter and reads the row tagged policy_denied. That row is the end of this flow. Durable audit is optional: with no writer configured, the default is to drop the record.