HAAS // HARNESS AS A SERVICE & MULTI-TENANT CLOUD ARCHITECTURE

Scale AI Agent Fleets with
Hardware-Isolated MicroVMs.

CloudHarness HaaS bridges self-hosted coding harnesses and elastic, multi-tenant cloud clusters. Architectural specification for executing untrusted coding agent workloads with sub-150ms cold boot target, hard block quotas, and double-entry metering.

⚠️
Architecture Preview & Design Specification: Commercial multi-tenant infrastructure is strictly gated behind ADR 0001 owner sign-off and empirical bare-metal KVM benchmarks. The private single-owner OSS installer remains the default production release.
Explore the Isolation Ladder ↓ Read Security Architecture ↗

The 3-Tier Tenant Isolation Ladder

Mutually distrustful tenants require hardware-level boundaries. CloudHarness specifies an explicit 3-tier isolation ladder to eliminate cross-tenant contamination.

Tier 1: OSS Community (Live)

Single Trusted Owner

Self-hosted deployment on your own Linux VPS for individual engineers and tightly-knit teams with mutual trust.

  • Rootful Docker Engine with seccomp and AppArmor
  • Loopback Ingress (127.0.0.1:3100) + Caddy TLS / Cloudflare Tunnel
  • Single-owner bearer token authentication
  • 1-Click single-command installation pipeline
Deploy OSS Server →
Tier 2: Commercial Beta (Private)

Dedicated Provider VM per Tenant

Hard hypervisor boundary provided by the cloud provider. Exactly one isolated cloud VPS provisioned per paying customer.

  • 1 Customer = 1 Cloud VPS (Hetzner / AWS / OVH)
  • Zero cross-customer container co-location
  • Outbound gRPC over mTLS Node Agent
  • Centralized control plane with capacity scheduling
Apply for Beta Access →
Tier 3: Commercial Pooled GA

Firecracker KVM MicroVMs

High-density hardware virtualization for public multi-tenant pooled clusters on bare-metal servers.

  • Dedicated Linux guest kernel booted in < 150ms
  • Jailer sandboxing with unique unprivileged UID/GID
  • Device Mapper (dm-thin) 20GB hard block quotas
  • FAIL-CLOSED Guarantee: Never falls back to Docker
Gated on Phase 04 KVM Verification

Outbound gRPC Node Agent (Design)

Proposed architecture: Worker nodes initiate outbound mTLS streams to the central control plane, eliminating all public inbound listening ports on execution nodes.

🛡️

Zero-Leak Node Enrollment

Tokens are structured as <id>.<secret> and verified via Argon2id with guarded database row consumption. Stdin ingestion (--token-stdin) or root-only 0600 file ingestion with descriptor-based O_NOFOLLOW checks prevents process table or bash history leaks.

⏱️

Watchdog Fence & Split-Brain Quarantine

If a worker node loses connection to the control plane for > 30 seconds, a local watchdog timer self-fences the node and pauses active containers. Reconnection requires monotonic lease_epoch revalidation before execution can resume.

Dedicated Tenant Scheduling

Database-enforced invariant: Dedicated tenant workspaces are routed strictly to nodes with matching nodes.tenant_id foreign keys. General tenant workloads are rejected from dedicated hardware enclaves.

Hybrid Metering & Double-Entry Ledger (Design)

Target model: Fair, fraud-resistant pricing combining duration tracking with tool execution request volume.

⏱️ 15-Minute Auto-Idle Sleep (Design Target)

Target specification: Workspaces with no incoming tool calls for 15 minutes are planned to automatically suspend, halting duration metering immediately and resuming in under 2 seconds.

💳 Double-Entry Financial Ledger (Design Architecture)

Planned design: Billing events execute an atomic PostgreSQL transaction (post_journal) ensuring balanced debits and credits with sub-cent micro-USD integer scale.

🔐 Raw-Body Webhook Verification (Design Architecture)

Planned design: Payment webhooks (Polar.sh and Stripe) are designed to validate HMAC signatures against raw byte buffers with timestamp tolerances (<300s) and composite key deduplication.

Data Residency & Incident Containment (Design)

Proposed security policy: Strict regional isolation with cryptographic erasure upon workspace termination.

🌍 Geographic Enclaves (Design Policy)

Proposed policy: Worker nodes are planned to be partitioned into regional silos (e.g. EU, US). Repository clones and ephemeral snapshots never cross geographical boundaries.

🔑 Ephemeral Key Destruction (Security Invariant)

Security invariant: Workspace termination wipes ephemeral AES-256 snapshot keys immediately. Unused storage blocks on thin-provisioned NVMe volumes are securely trimmed.

🚨 Automated Incident Quarantine (Security Policy)

Incident runbook: Syscall anomalies or jailer seccomp violations trigger instant node cordoning, certificate revocation via CRL, and tamper-evident audit logging.

🔐 Encrypted Forensic Capture & Default Erasure

Under explicit security authorization, memory state is captured directly into an access-controlled, asymmetric encrypted container (AES-256-GCM + ephemeral key pair) with mandatory 30-day retention. By default, unmanaged rogue microVMs are terminated immediately with complete cryptographic key erasure and zero plaintext memory preservation (ADR 0001 Design).

Ready to Scale Dedicated AI Coding Infrastructure?

Join our invite-only Dedicated Commercial Beta for engineering teams deploying high-concurrency coding agents.

GitHub Community Issues ↗ Deploy Self-Hosted OSS →