The platform

What Rokha AI ships

A software-as-a-service execution layer for AI agents. Humans get a workspace — chat, a visual builder, scheduling. Agents get the same platform through machine-legible doors: an OpenAPI surface, an MCP server, and A2A endpoints, designed so another agent can go from discovering a capability to a real invocation without a human click. Both fronts sit on one runtime.

• The registry — every major agent skills, MCP servers and agents indexed from nine sources, searchable by keyword, fuzzy match and embeddings.
• The runtime — tool-shaped work dispatches to a throwaway, isolated sandbox; prompt-shaped work runs agentically. Every run is metered, budgeted server-side, and recorded.
• Traces — the runtime's receipts. Every execution writes a structured record of what ran, what it touched, and what it returned — verifiable output, not claimed output.
• The SDK and CLI — typed clients, a published OpenAPI schema, and a local terminal agent, so the platform composes into other stacks.

Architecture

Four Rust services behind one router, each independently deployable, with Postgres underneath and every frontend and agent request passing through the same trust boundary.

clients (web · SDK · MCP · A2A)
        │
     erebus — router, auth, billing, trust boundary
        │
  ┌─────┼──────────┐
agents  harnesses  protocols     ── Rust services
        │
  sandbox runtime                ── isolated per-run containers:
  (AWS Fargate, x86-64)             no secrets, egress-only,
                                    fail-closed budgets

The sandbox is the platform's security core: each run gets a fresh container with no credentials inside, least-privilege egress-only networking, and hard resource and spend ceilings that fail closed. Untrusted third-party tools execute there — never in the services.

Silicon

Rokha is built on Intel® architecture: containerized Rust on x86-64 cloud compute (AWS ECS and Fargate), with no GPU dependency in the serving path — and the platform's AI inference layer is Intel-based, served from a dedicated Intel® Xeon® node running OpenVINO™. CPU-class inference is a deliberate architectural fit for an agent runtime whose work is orchestration, tool execution and retrieval.

• Today — the full platform on x86-64 cloud infrastructure, and a dedicated Intel® Xeon® inference node running an OpenVINO™ cross-encoder that re-scores registry search for Studio callers — a real production inference workload on Intel silicon, and the node stamps every response with its own CPU reading.
• In build — the ro local terminal agent packaged for Intel-based AI PCs, keeping light inference on-device; broader OpenVINO™ coverage (embeddings, classification) as the workloads grow.

We label the line between shipped and planned on purpose — this platform's standing rule is that nothing pretends to run. Hardware claims here follow the same rule.

Security posture

• Hostile-caller by default: every door authenticates, every id is ownership-checked.
• Money, quota and authority checks fail closed — an unreadable guard refuses.
• Sandboxed execution carries no secrets and reaches the network egress-only.
• The model routes; it never authorizes — real permission checks sit behind every effect.
• Full audit trail: traces on every run, receipts on every payment.

Partner with Rokha AI

Rokha AI is a U.S. software company (LLC) operating rokha.ai as a production service. We're an official partner in three programs today: the Intel® Partner Alliance (Independent Software Vendor, with a storefront on the Intel® Partner Showcase), IBM Partner Plus (Registered tier) and the Microsoft AI Cloud Partner Program (Independent Software Vendor; a Microsoft Marketplace listing is in preparation). NVIDIA is next in focus. We're interested in hardware and platform partnerships — ISV programs, co-engineering on CPU-class inference, and edge/AI-PC distribution for the local agent.

Ten slides on the product as it runs today: what Rokha is, the numbers, the team.