Tool categories, by layer.
The tool categories that fill each layer of the stack, with 96 illustrative examples across 31 categories. The category is the substance; the brands are examples, each with its licence and a checked link. See how they fit together in the stack.
The five layers
Each category is keyed to the layer it serves. The colour ticks throughout follow this legend.
- L1 Govern-as-Code
- L2 Inventory & Transparency
- L3 Evals & Red Teaming as Evidence
- L4 Runtime Controls & Observability
- L5 Assurance & Continuous Compliance
The catalogue
Every category says what it does for governance and where the Body of Knowledge treats it. Filter by layer or by licence; the counts update as you choose.
-
Policy engines
Evaluate a rule written as code against an input and return a structured allow or deny verdict.
-
Keeps policy engines in sync with policy and data sources.
Chapter 04, layer 01 → Obligations this answers → Patterns for layer 01 →
-
Policy testing and authoring
Write and test policies before they gate anything: unit tests for rules, run in CI against configuration files.
-
Hosted at openpolicyagent.org.
Chapter 04, layer 01 → Obligations this answers → Patterns for layer 01 →
-
Admission control
Enforce policy at the moment a workload is admitted to a cluster, so an unregistered or unsigned deployment never starts.
Chapter 04, layer 01 → Obligations this answers → Patterns for layer 01 →
-
Machine-readable policy artefacts (proposed)
Proposed formats that state an agent's rules as data an enforcement point and an auditor can both read.
-
A research proposal (arXiv 2510.24383).
Chapter 04, layer 01 → Obligations this answers → Patterns for layer 01 →
-
-
Registries and governance suites
Hold the inventory of AI systems with an owner, a scope and a status per entry.
-
AI Control Tower product page.
Chapter 04, layer 02 → Obligations this answers → Patterns for layer 02 →
-
-
Agent-discovery tools
Find the agents and AI features already running, including the ones nobody registered.
Chapter 04, layer 02 → Obligations this answers → Patterns for layer 02 →
-
AIBOM formats and generators
Record the models, datasets, weights and dependencies of an AI system as a machine-readable bill of materials.
Chapter 04, layer 02 → Obligations this answers → Patterns for layer 02 →
-
Model and data card tooling
Write the transparency documents (model cards, dataset cards) as structured metadata beside the artefact.
-
Card tooling ships in the huggingface_hub library.
Chapter 14, model cards and datasheets → Obligations this answers → Patterns for layer 02 →
-
-
FRIA/DPIA tooling
Run a structured impact assessment and keep its result linked to the registry entry it concerns.
Chapter 14, impact assessments → Obligations this answers → Patterns for layer 02 →
-
PII detection and redaction
Find and mask personal data in text, images and tables before it enters a prompt, a log or a training set.
Chapter 19, privacy and AI → Obligations this answers → Patterns for layer 02 →
-
Data and experiment versioning
Version the data, code, parameters and lineage behind each model so a result can be reproduced and traced back.
-
An open standard for lineage events.
Chapter 14, linked versioning → Obligations this answers → Patterns for layer 02 →
-
Model signing and artefact scanning
Sign model artefacts, attest how they were built and scan them for unsafe serialisation before they load.
-
A framework of supply-chain provenance levels.
Chapter 14, linked versioning → Obligations this answers → Patterns for layer 02 →
-
Evaluation frameworks
Run evals as code, versioned with the model, so a regression can fail the build.
Chapter 04, layer 03 → Obligations this answers → Patterns for layer 03 →
-
Adversarial and vulnerability probes
Probe a model or an agent for jailbreaks, prompt injection and tool misuse as repeatable tests.
Chapter 04, layer 03 → Obligations this answers → Patterns for layer 03 →
-
Safety and red-team benchmarks
Shared test sets and harnesses that make one model's safety result comparable with another's.
-
The public repository is published under Apache-2.0.
Chapter 04, layer 03 → Obligations this answers → Patterns for layer 03 →
-
-
Retrieval-augmented quality
Score a retrieval-augmented system on grounding and relevance, not only on fluent answers.
Chapter 04, layer 03 → Obligations this answers → Patterns for layer 03 →
-
Data validation and quality
Test a dataset against declared expectations (schema, ranges, completeness, distributions) and fail the pipeline when it breaks them.
Chapter 14, data quality → Obligations this answers → Patterns for layer 03 →
-
Fairness toolkits
Compute disparity metrics across groups and apply mitigation before, during or after training.
Chapter 16, group fairness metrics → Obligations this answers → Patterns for layer 03 →
-
Explainability libraries
Produce feature attributions, surrogate models and counterfactuals, then test that the explanation is faithful.
-
Counterfactual explanations.
Chapter 16, explanation techniques → Obligations this answers → Patterns for layer 03 →
-
Practice grounds
Hands-on challenges for learning how prompt injection works before testing your own system.
-
Redirects to Lakera's challenge hub (as of 2026-09-24).
Chapter 04, layer 03 → Obligations this answers → Patterns for layer 03 →
-
-
Guardrail frameworks
Filter inputs and outputs and mediate tool calls at the enforcement point, on the live call.
-
Open weights under a custom licence.
Chapter 04, layer 04 → Obligations this answers → Patterns for layer 04 →
-
Observability
Trace every model and tool call so agent behaviour becomes a control signal and a record.
-
Core under MIT; enterprise directories under a commercial licence.
Chapter 04, layer 04 → Obligations this answers → Patterns for layer 04 →
-
-
ML and LLM monitoring and drift
Watch inputs, outputs and performance in production and raise a signal when they drift from what was approved.
Chapter 15, drift → Obligations this answers → Patterns for layer 04 →
-
MCP / tool-call security
Inspect and scan the servers and tools an agent connects to, and guard what flows back into its context.
-
The MCP project is moving its code from MIT to Apache-2.0.
-
Formerly mcp-scan (Invariant Labs); the old repository redirects here.
Chapter 04, layer 04 → Obligations this answers → Patterns for layer 04 →
-
-
Progressive delivery and feature flags
Expose a change to a small share of traffic first and roll it back automatically when a guard metric fails.
-
A vendor-neutral feature-flag API.
Chapter 15, progressive delivery → Obligations this answers → Patterns for layer 04 →
-
Kill switch / circuit breaker
Stop one agent or one feature quickly and safely, without taking the rest of the system down.
-
Feature flags, including kill-switch flags.
-
Gateway-level circuit breaking.
-
In-process circuit breakers.
Chapter 04, layer 04 → Obligations this answers → Patterns for layer 04 →
-
-
Agent workload identity
Give each agent its own workload identity and scope, so its access can be bounded, logged and revoked.
-
Launch announcement.
Chapter 04, layer 04 → Obligations this answers → Patterns for layer 04 →
-
Evidence format
State controls, implementations and assessment results in one machine-readable format an auditor can query.
Chapter 04, layer 05 → Obligations this answers → Patterns for layer 05 →
-
OSCAL tooling (open source)
Author, validate, convert and resolve OSCAL content in a pipeline instead of by hand.
Chapter 04, layer 05 → Obligations this answers → Patterns for layer 05 →
-
GRC and AI-governance suites
Aggregate evidence from the lower layers and present it against frameworks for review.
Chapter 04, layer 05 → Obligations this answers → Patterns for layer 05 →
-
Curated indexes
Across layersMaintained lists that go wider than this catalogue; start here when a category needs more options.
No tool in the catalogue matches both filters.