Ten studios behind one API
Chat, agentic chat, agents, retrieval, vision, video, speech, document AI, training and workflow are separate surfaces over a shared platform. An application integrates once and reaches all of them.
AI infrastructure · live
CoralMeshAI
A multi-tenant AI platform: ten studios, a workflow engine, GPU resource management and an OS-level desktop agent on one token meter.
Overview
An enterprise platform consolidating the AI surface an organisation would otherwise assemble from a dozen vendors. Ten studios cover conversational chat, agentic chat, autonomous agents, retrieval, vision, video, speech, document intelligence, model training and workflow composition. Underneath, a NestJS and PostgreSQL backend handles multi-tenancy, authentication, conversation storage, billing, usage analytics and GPU resource allocation. A separate desktop agent extends execution to the operating system itself, running sandboxed native tools under enforced permissions with audit logging. Everything is metered at request, session, user and tenant level, and the platform ships as Docker images ready for Kubernetes, so it can run on hardware the organisation already owns.
Specifications
- Audience
- AI infrastructure
- Platforms
- Self-hosted · Web
- Status
- live
- Built with
- OpenAI-compatible API
Capabilities
A desktop agent with real operating-system access
The agent executes native tools on macOS and Windows inside a sandbox, under policy enforcement with explicit user consent. Credentials are held in the OS keychain or credential manager, execution is deterministic, and every action is written to an audit log.
Voice-triggered, workflow-driven execution
Hotword detection starts a session hands-free, and workflows defined in the cloud can target desktop execution — a workflow authored centrally runs its operating-system steps on the operator's own machine.
Multi-tenancy through the whole stack
Tenant isolation runs from authentication through conversation storage to billing. Platform administration, tenant administration and end-user surfaces are separate applications over the same core.
Metering at four levels
Tokens are counted per request, per session, per user and per tenant. Usage analytics and billing read the same ledger, so the dashboard and the invoice cannot disagree.
GPU resource management in the console
Model serving is scheduled against the GPUs registered to the platform, with allocation and utilisation visible in the console rather than hidden behind a provider abstraction.
Embeddable and self-hosted
An embed surface exposes the assistant inside another product. The platform itself is packaged for Docker and Kubernetes, so no request needs to leave the organisation's network.
Others in this line
Portfolio products serving the same audience.
Engineered independently. Priced transparently.
No per-seat licensing, no vendor lock-in, no unpredictable invoicing. Our teams are available to scope an evaluation for any product in the portfolio.