Govern AI systems with evidence, policy and operational control.
KAVACH is an open governance and control plane for AI agents, workflows and decisions. It provides evaluation, policy enforcement, execution control, auditability and deterministic governance across REST APIs and MCP tools.
Architecture
Layered Control Plane for AI Governance
KAVACH separates experience, control, governance, execution and persistence into explicit planes with unidirectional data flow.
Capabilities
Core Governance Building Blocks
Each subsystem is independently deployable and exposed through REST APIs and MCP tools.
Job Execution Control Plane
Record and execute long-running governance work — evaluations, experiments, replay, and drift analysis — with durable lifecycle tracking.
Evaluation Plane
Turn workflow outputs into evaluation results using pluggable providers. Providers sit behind contracts, not platform boundaries.
Governance Ontology
A semantic graph describing relationships between governed assets — prompts, models, datasets, decisions, and evidence.
Governance Decisions
Explainable, auditable outcomes generated from evidence and policy. Trace lineage from decision back to every supporting artifact.
MCP Server
Expose governed read and controlled-write operations to agents through auditable MCP tools, each delegating to one REST endpoint.
Settings Control Plane
Typed configuration with SYSTEM, ORGANIZATION and PROJECT scopes. Compare-and-set updates with versioned audit trails.
Audit & Replay
Capture workflow and node-level execution records. Reconstruct historical executions for investigation, debugging and governance review.
Drift Analysis
Compare evaluations over time, calculate drift and produce governance-oriented artifacts from historical data.
Replaceable Persistence
SQLite for local development, PostgreSQL for production, Snowflake for analytics. Repository contracts keep storage interchangeable.
Lifecycle
How KAVACH works
A deterministic loop from observation to immutable audit.
Observe
Capture execution, evidence, model, prompt and policy context.
Evaluate
Run quality, safety, compliance and governance checks.
Decide
Produce deterministic governance decisions with reasoning and evidence.
Audit
Store immutable execution and decision history for replay and investigation.
Principles
Engineering Principles
KAVACH is built to be operated in production. These constraints shape every interface.
Control Plane Ownership
Kavach owns governance state and decisions. It tracks assets, results, comparisons, rankings and recommendations — but never the live AI runtime or deployment system.
Explicit Contracts
Every boundary is visible in code. Repository interfaces, provider interfaces and framework-neutral APIs are all explicit contracts. Hidden cross-system contracts are avoided.
Immutable Artifacts
Governed records are immutable where possible. Mutable control flows create new records or versions instead of mutating historical facts.
Version Everything
Prompts, models, datasets and experiment-facing artifacts are versioned. Versioning is required for reproducibility, comparison, auditability and replay.
Replayability
If a governance decision cannot be reconstructed later, it is weak evidence. Replayability affects storage, versioning, immutability and API design.
Storage Independence
The platform does not depend on SQLite as a core architectural choice. Repository contracts preserve the ability to add PostgreSQL, Snowflake or other stores.
Provider Independence
Evaluation logic remains behind provider interfaces. TruLens is the current implementation, but the architecture should not require provider-specific knowledge.
Failure Isolation
Failures stop at the narrowest useful boundary. A provider failure should not silently mutate registry state. A transport adapter should not rerun ranking logic.
Interfaces
Evaluate Governance Decisions
The same governance operation is available over REST and as an auditable MCP tool. Every call is correlated and recorded.
POST /api/v1/decisions/evaluate
{
"subject_type": "agent_execution",
"subject_id": "execution-2048",
"policy_id": "production-ai-policy",
"evidence_ids": [
"eval-result-981",
"audit-event-477"
]
}
// 200 OK
{
"decision_id": "decision-991",
"outcome": "APPROVED",
"policy_version": "3.2.0",
"reason": "All mandatory controls satisfied"
}Releases