Enterprise Overview¶

KruxOS provides the governance, security, and auditability that production AI agent deployments require.

The challenge¶

Deploying AI agents in production creates unique risks:

Uncontrolled access — agents with shell access can read any file, run any command, access any network service
No audit trail — when something goes wrong, there's no record of what the agent did or why
No governance — no way to enforce policies, require approvals, or rate-limit operations
Secret exposure — API keys and credentials passed as environment variables are visible to agents
No isolation — multiple agents share the same filesystem, processes, and network

How KruxOS solves this¶

Risk	KruxOS Solution	Implementation
Uncontrolled access	Per-agent sandboxing	Linux namespaces + cgroup v2 + seccomp + nftables (Landlock MAC adds in v0.0.3)
No audit trail	Hash-chained audit logs	Append-only CBOR files with SHA-256 chain, SQLite index
No governance	Deterministic policy engine	YAML rules compiled to evaluation tree, 4 permission tiers
Secret exposure	Use-not-read vault	AES-256-GCM encrypted storage, agents never see raw values
No isolation	Per-agent resource limits	cgroup v2 CPU/memory/IO/PID limits per agent

Architecture¶

graph TB
    subgraph "Agent Layer"
        A1[Agent 1<br/>Claude]
        A2[Agent 2<br/>GPT-4o]
        A3[Agent 3<br/>Gemini]
    end

    subgraph "KruxOS"
        GW[Gateway<br/>MCP / JSON-RPC<br/>:7700]
        PE[Policy Engine<br/>YAML rules]
        CR[Capability Registry<br/>89 capabilities]
        SB[Sandbox<br/>per-agent isolation]
        VT[Vault<br/>encrypted secrets]
        AL[Audit Log<br/>hash-chained]
        SP[Service Proxy<br/>read-replica + write buffer]
        ST[State System<br/>3-tier persistence]
        CM[Agent Comms<br/>message broker]
    end

    subgraph "External"
        Gmail[Gmail API]
        Other[Other Services]
    end

    subgraph "Supervision"
        DB[Dashboard :7800]
        CLI[kruxos CLI]
    end

    A1 & A2 & A3 --> GW
    GW --> PE
    GW --> CR
    GW --> SB
    GW --> VT
    GW --> AL
    CR --> SP
    SP --> Gmail & Other
    GW --> ST
    GW --> CM
    DB & CLI -->|Supervision :7701| GW

Key differentiators¶

Model-agnostic governance¶

KruxOS works with any AI model — Claude, GPT, Gemini, Llama, or any custom model. The governance layer is model-independent. Policies, sandboxing, and audit apply uniformly regardless of which model is driving the agent.

Zero-trust architecture¶

Every agent operation goes through the same pipeline: authenticate → evaluate policy → sandbox → execute → audit. There are no backdoors, no admin-mode bypasses, and no way for an agent to skip the policy check.

Deterministic policy evaluation¶

Policy decisions are reproducible. Given the same agent, capability, and parameters, the policy engine always returns the same result. No LLM is involved in governance decisions. Every decision can be explained, replayed, and audited.

Production-ready safety¶

The Service Proxy prevents agents from causing damage to external services through:

Read-replicas — read operations never touch the external service
Write buffers — outbound operations are delayed, giving humans time to cancel
Batch protection — bulk operations automatically escalate to human approval
Soft-delete — destructive operations preserve originals for 24-hour recovery

Documentation¶

Page	What you'll learn
Security Model	Sandboxing, vault, audit chain, policy hierarchy
Architecture	System architecture with data flow diagrams
Compliance	SOC2, ISO27001 readiness
Deployment Guide	Production deployment best practices
Service Proxy	External service safety model
Benchmarks	Performance and token efficiency results
Comparison	vs other agent deployment approaches
Pricing	Community vs Enterprise editions
Contact	Talk to the KruxOS team