Your agents do the work.
We make sure they follow the rules.

Five lines to a pre-execution enforcement gate that writes to your existing Postgres. Enforcement gates for every action routed through the SDK — budget caps, scope checks, loop detection, human-in-the-loop approvals, agent presence monitoring, and a tamper-evident audit trail that compliance can verify themselves. See the threat model for what the SDK does and does not protect against.

MIT licensed. Free and open source.

import os
from codeatelier_governance import GovernanceSDK, AuditEvent

async with GovernanceSDK(database_url=os.environ["DATABASE_URL"]) as sdk:
    await sdk.audit.log(AuditEvent(agent_id="my-agent", kind="hello"))

from codeatelier_governance import ScopePolicy, BudgetPolicy

sdk.scope.register(ScopePolicy(
    agent_id="billing-agent",
    allowed_tools=frozenset({"read_invoice", "send_email"}),
    hidden_tools=frozenset({"delete_all"}),  # hidden from LLM context
))

sdk.cost.register(BudgetPolicy(
    agent_id="billing-agent",
    per_session_usd=0.50,
    per_session_seconds=300,  # 5-minute session time limit
    per_agent_usd_daily=10.00,
))

# Auto-compute cost from model name — no manual usd= needed:
await sdk.cost.track_usage("billing-agent", session_id,
    model="gpt-4o", input_tokens=1000, output_tokens=500)

import os
from codeatelier_governance import GovernanceSDKSync, AuditEvent

with GovernanceSDKSync(database_url=os.environ["DATABASE_URL"]) as sdk:
    sdk.audit.log(AuditEvent(agent_id="my-agent", kind="tool.call"))
    sdk.scope.check("my-agent", tool="send_email")
    sdk.cost.check_or_raise("my-agent", session_id)

Most tools tell you what your agent did. After the damage is done.

LangSmith, Langfuse, and Helicone are observability platforms — they log what happened, after the fact. That does not stop a runaway agent from burning $600 overnight, calling a tool it should never touch, or modifying a patient record without approval. The Governance SDK is different: it gates decisions before the LLM call fires.

Agent frameworks — including Microsoft Agent Framework, which now ships DevUI, Aspire, and first-class HITL — solve the build-an-agent problem. Our differentiation is narrower and earlier: HMAC-chained tamper-evident audit + just-Postgres + scope, budget, and HITL gates in one primitive, callable from any framework, with evidence a regulator can verify.

Enforcement modules. One substrate.

Core Enforcement

Extended Modules

Fast enough to gate every call.

300+ tests passing. MIT licensed.

Enforcement, not just tracing.