Production Status: Operational
EU AI Act Annex III enforcement target:
December 2, 2027 May 2026 Digital Omnibus agreement.
Subject to formal adoption.

OPEN SOURCE / STDLIB-ONLY / APACHE 2.0

Governed systems for agentic work.

HUMMBL builds the control plane for AI-native engineering teams: delegation tokens, governance buses, kill switches, circuit breakers, audit receipts, and public proof surfaces that run in customer-controlled environments.

Build with the Library → Inspect Evidence
15,600+ Aggregate tests (incl. internal)
7 Governance primitives (of 34 in package)
0 Core runtime dependencies
120 Mental models
hummbl-governance
$ pip install hummbl-governance
Published on PyPI · Open Source on GitHub

The Governance Tuple

Every governed AI decision produces a triple: CONTRACT (what was authorized), DCT (who authorized it, cryptographically signed), and EVIDENCE (what actually happened). This is the atomic record that makes AI governance auditable.

CONTRACT
Permitted operations, resources, caveats, expiry
×
DCT
HMAC-signed delegation binding issuer → subject
×
EVIDENCE
Actual operations, resources accessed, cost, output
governance_tuple.py 
from hummbl_governance import DelegationTokenManager, BusWriter, CostGovernor

# CONTRACT: define what the agent may do
contract = {
    "ops_allowed": ["read", "summarize"],
    "resources": ["docs/*"],
    "expiry": "2026-12-31T23:59:59Z"
}

# DCT: cryptographically bind the delegation
dtm = DelegationTokenManager()
token = dtm.issue(
    issuer="orchestrator",
    subject="summarizer-agent",
    operations=contract["ops_allowed"],
    resources=contract["resources"]
)

# EVIDENCE: audit what actually happened
bus = BusWriter("governance.jsonl")
bus.append(agent="summarizer-agent", action="read", resource="docs/report.md")
bus.append(agent="summarizer-agent", action="summarize", resource="docs/report.md")
View Architecture → Read the Paper → OWASP 10/10 Coverage →

Cognitive Layer: Base120 Mental Models

Base120 is the secondary governed reasoning substrate that operates over runtime governance primitives. It keeps model selection, problem framing, and reasoning transformations inspectable without turning the homepage into a package-status ledger.

P
Perspective
Frame and anchor viewpoint. Change how you see the problem.
IN
Inversion
Flip to see what avoiding failure reveals. Think backwards.
CO
Composition
Combine for emergent properties. Build complexity from simplicity.
DE
Decomposition
Break complexity down. Understand the parts.
RE
Recursion
Apply patterns at multiple scales. Iterate and refine.
SY
Systems
Understand governing rules. See the meta-patterns.

Try It Now

Describe a problem you're facing and get recommended mental models instantly. No signup required.

Recommended Models

Explore All 120 Models →

Use Cases

HUMMBL is for teams putting agents into production workflows where authority, containment, evidence, compliance, and release risk must be explicit.

Agent Fleet Authority

Scope what every agent may do with HMAC-signed delegation tokens: operations, resources, caveats, expiry. No ambient permissions.

Runtime Containment

Kill switches, circuit breakers, and cost governors that stop a misbehaving agent in-flight, not in next sprint's postmortem.

Audit-Ready Evidence

Append-only receipts for every governed decision. Answer who authorized this and what actually happened in minutes, not a forensics engagement.

EU AI Act Readiness

Map runtime evidence to Annex III obligations ahead of the December 2027 target. Compliance generated by the system, not written about it.

Governance Gates in CI/CD

Block merges and deploys that violate declared contracts. The same gates that validate this site validate your pipeline.

Procurement-Grade Proof

Public trust surfaces with source-linked claims: hand security reviewers a validation page instead of a questionnaire thread.