KDCube - safer way to build AI for companies

Deploy AI agents with enforceable runtime controls: block unsafe actions, contain spend, and keep company data inside your infrastructure.

Business-first summary lives here. Deep technical behavior is documented in Docs and control details in Security.

  • Unsafe tool actions are blocked before execution, not after incident review
  • Hard spend and rate limits are enforced outside prompt logic
  • Tenant boundaries are enforced across gateway, storage, and data access paths

For architecture and controls by layer, use docs.html and security.html.

Star on GitHub MIT License Self-Hosted Open Source

Execution Safety

Agent-generated code executes in isolation. Privileged actions are mediated by runtime controls.

Cost Governance

Set hard limits for users, projects, and organizations to prevent runaway usage.

Data Boundaries

Tenant separation is enforced across request routing, storage, and accounting.

Self-Hosted Control

Run on your infrastructure with audit visibility and no mandatory vendor data path.

Why sandboxing alone is not enough

Compute isolation is necessary, but companies also need policy, spend, and tenant controls before actions are executed.

Infrastructure sandboxing

Good at isolating compute and limiting process-level access.

Not enough for business controls like approval policy, tenant scope, and cost containment.

What KDCube adds

  • Action mediation: privileged operations pass through controlled runtime boundaries.
  • Cost governance: spend and rate controls are enforced independently of prompt output.
  • Tenant enforcement: request and data access paths are scoped by tenant and project.
  • Auditability: decisions are traceable on infrastructure you control.

The enterprise risk is policy failure, not only process escape. KDCube focuses on preventing unsafe or out-of-scope actions before they execute.

Sandboxing code is not enough.

AI agents don't just run code. They write emails. They process refunds. They query customer records. They call third-party APIs with real consequences.

Infrastructure sandboxes isolate the compute. Guardrails wrappers filter the output. Observability tools log what happened. None of them stop a wrong action before it executes.

The enterprise risk is not malicious code. It is an agent that does something it was never supposed to do — and nobody blocked it in time.

Infra Sandbox

What it does: Isolates execution environment

What it misses: Does not know what the agent intended to do

Guardrails Wrapper

What it does: Filters output after generation

What it misses: Does not intercept tool calls or API actions

Observability Tool

What it does: Logs what happened

What it misses: Logs are read after the fact — damage already done

The gap is semantic. KDCube closes it.

What is a semantic sandbox?

From infrastructure sandbox → semantic sandbox.

Traditional Sandbox

Isolates where code runs. Controls the compute environment. Does not understand what the agent is trying to do.

KDCube Semantic Layer

Enforces budget, rate, tenant, and execution-safety boundaries before execution. Declarative action-level policy DSL is roadmap.

KDCube intercepts agent execution at runtime and applies configured constraints (budget, rate, tenant, and isolation) before calls are admitted. Deterministic per-action allow/deny across arbitrary workflows is on the roadmap.

How KDCube works

Admit, enforce, execute, and audit. Implementation detail is in Docs.

Subprocess-isolated code execution

  • Isolated subprocess within Docker Compose deployment
  • No external network access by default
  • Sensitive env vars stripped; minimal filtered environment; workdir filesystem only

✅ Available

Budget controls and rate limiting

  • Per-user, per-project, and per-org spending caps with multi-tier accounting rollups
  • Reservation/commit economics checks at admission with hard-limit enforcement
  • Request frequency and token throughput constraints

✅ Available

Multi-tenant isolation patterns

  • Tenant boundary validated on every request
  • Cross-tenant DB access blocked at the gateway layer
  • Knowledge base scoping per tenant (validate KB isolation boundaries during setup)

✅ Available

Streaming runtime with tool orchestration

  • REST, SSE (Server-Sent Events), Socket.IO
  • Token-by-token streaming responses
  • Composable Skills and tool namespaces (local + MCP) via dynamic bundles

✅ Available

Decision logging (request/response)

  • Every allow/block decision timestamped
  • Self-hosted on your infrastructure
  • Control plane monitoring dashboard included

✅ Available

Policy DSL & deterministic enforcement Roadmap

  • Declarative constraint definitions per agent role
  • Verifiable allow/deny before any tool call fires
  • Workflow invariants and cross-agent approval gates

🔮 Roadmap

Where semantic sandboxing matters

Enterprise scenarios where "block before execution" is not optional.

Refund Authorization

An agent processes customer refund requests. KDCube enforces a maximum refund per transaction. Anything above requires a human approval step. The agent cannot bypass the limit — not through prompt manipulation.

$500 hard cap enforced

CRM Access Boundaries

An agent assists sales reps. A prompt injection attempt asks the agent to retrieve data from a different tenant. KDCube enforces tenant boundaries at the gateway layer — cross-tenant data access is blocked before it reaches the database. Validate KB isolation boundaries during setup.

Zero cross-tenant leaks

Approval Workflows

An agent manages vendor onboarding. It can draft contracts but must not send them without a human sign-off. KDCube enforces current admission/economics/runtime boundaries today; explicit workflow-step invariant gating is a roadmap capability.

Enforced approval gates Roadmap

Data Boundary Enforcement

An agent processes medical records. KDCube restricts outbound API calls to an approved allowlist. Calls to unapproved endpoints are blocked. The execution environment has no external network access by default.

Allowlist-only outbound calls

Cost Containment

An agent runs complex analytical tasks. Left unconstrained, a single poorly-scoped prompt can trigger thousands of dollars in LLM token consumption. KDCube enforces per-user, per-project, and per-org spending caps. Hard limits cannot be exceeded.

Potential 60–80% infra savings in self-hosted scenarios*

Secure Code Execution

An agent generates and runs Python scripts. KDCube runs the code in an isolated subprocess with no external network access and no environment variable access. No code run in the sandbox can reach production systems directly — all outbound tool calls route through the trusted supervisor.

Zero network egress by default

*Estimate range depends on workload profile, model mix, traffic shape, and infrastructure/operations choices.

Why KDCube

Not a framework. Not a guardrails wrapper. Not observability.

Block before execution — not after

Most tools observe or filter. KDCube intercepts. The enforcement decision happens before any external system is called. Wrong actions do not execute.

Semantic constraints, not just compute limits

Rate limits and memory caps are infrastructure. KDCube enforces what the agent is allowed to do — action scope, data access, spend authority — not just how much CPU it uses.

Self-hosted. Your infrastructure, your policy.

KDCube runs on your infrastructure. No vendor has access to your agent interactions, your policies, or your data. MIT licensed. No lock-in.

Open-source and auditable

The enforcement layer is not a black box. Every constraint, every decision, every blocked action is logged and inspectable. You own the audit trail.

Open-source. Self-hosted. Zero vendor lock-in.

Fast evaluation path

Use the all-in-one deployment to validate fit quickly, then move to managed infra patterns for production.

  • Typical first environment in under an hour
  • Clear migration path to production topology
  • Full setup steps in Deployment Model

What you get

  • Business-safe default posture for spend, tenant separation, and controlled execution
  • Lower platform lock-in risk with self-hosted deployment choices
  • Your data stays on your infrastructure by default

KDCube is MIT licensed. The full source is on GitHub. Deploy on your own infrastructure — a single VPS, Docker Compose, or a Kubernetes cluster — and you control everything.

No usage-based fees. No per-seat pricing. No vendor access to your agent traffic. Infrastructure costs vary by cloud provider and configuration; example ranges of $50–$2,000/month are based on typical VPS/cloud pricing. The runtime itself is free.

Where we are headed

The current runtime already provides economic controls, isolation, and Observability signals (monitoring endpoints, queue pressure, circuit-breaker state) used for scaling decisions. The following items are additional roadmap capabilities and are not yet shipped.

Policy DSL Roadmap

Declare agent permissions in a human-readable policy language. Define what actions, data scopes, and spend limits are permitted per agent role.

Deterministic Enforcement Engine Roadmap

Pre-execution evaluation that produces a verifiable allow/deny decision before any tool call fires, with no probabilistic components.

Workflow Invariants Roadmap

Declare required steps in a workflow. Prevent agents from skipping approval gates, notification steps, or compliance checkpoints.

Cross-Agent Approval Gates Roadmap

Require a second agent, human-in-the-loop confirmation, or external approval before high-impact actions execute.

Follow progress: github.com/elenaviter/kdcube-ai-app

KDCube vs. the alternatives

KDCube integrates with common agent frameworks and tool APIs. It is complementary to, not a replacement for, your orchestration layer.

Comparison of AI agent tooling categories: what each tool does and what it does not do
Category What it does What it does NOT do
Infrastructure sandbox
(Docker, gVisor)
Isolates compute environment Does not know what the agent intends to do; no semantic constraints
Guardrails wrapper
(NeMo Guardrails, etc.)
Filters LLM output text Does not intercept tool calls or API actions; fires after generation
Observability tool
(LangSmith, Helicone)
Logs and traces agent behavior Read-only; logs damage after it occurs; does not block
Citations / provenance layer Maintains source pools and citations so outputs can reference supporting sources Does not enforce admission controls, budgets, or tenant isolation by itself
Skills extension layer Adds namespaced Skills and MCP-backed tools to bundle workflows Does not replace Gateway / Admission checks or economics enforcement
Agent framework
(LangGraph, LangChain, CrewAI)
Orchestrates agent logic and tool calling Provides no enforcement layer; delegates execution control to the agent
KDCube — Semantic Sandbox Intercepts tool calls and API actions; enforces budget caps, rate limits, and tenant boundaries before execution (full semantic workflow constraints are on the roadmap) Not an LLM proxy, not a framework, not a log aggregator

Build company AI with safer defaults.

Use KDCube to deploy AI with enforceable controls, then scale with docs-backed architecture patterns.

MIT Licensed · Self-Hosted · Open Source