The AI layer business tools are built on.
Agents grounded in your knowledge, a model gateway with failover, embeddings and decision models, voice, and a harness where agents do real work on code and documents. It is the layer under products like Trelvia AI.
Access is invite-only — sign in with the credentials your admin provisioned.
How a request flows
Multi-agent RAG
Spin up any number of chatbots, each with its own system prompt and knowledge. Hybrid keyword + semantic search grounds every answer.
OKF libraries from git
Point an agent at any git repo holding an OKF bundle — markdown with YAML frontmatter. The platform keeps a checkout and the agent browses it like a folder tree, reading whole documents instead of chunks. Your repo stays the source of truth.
Tools that run in your frontend
Register the actions your app can perform, and the agent calls them: it hands the tool call back to your client, which runs it on the device and returns the result. Tune when a tool fires from the dashboard, without shipping a new app build.
AI-managed websites
Website owners change their site by chatting — on the dashboard or on Telegram. The agent reads the site, asks when a request is unclear, edits the code, commits and pushes; the site's own build publishes it.
An agent harness for real work
Agents that need files and programs work on our own servers: an isolated workspace per project with file, shell, git and job tools, reachable only through the platform. Reasoning and prompt caching are kept intact across long tasks.
Voice
A whole spoken turn over one WebSocket: speech-to-text, the agent's answer, and speech back sentence by sentence, interruptible mid-reply. Speech-to-text is also a plain API endpoint.
One gateway, two formats
Point existing OpenAI or Anthropic SDKs at the gateway. Alias models, route them to any upstream, and retarget every connected client with a one-row change.
Model failover
Give any chat model up to two backups. When it fails before answering — a network error, a rate limit, an outage — the same request moves to the next one, and agents keep working.
Embeddings and decisions
Embedding models at /v1/embeddings, and typed-decision models at /v1/decisions that answer questions with calibrated probabilities instead of prose — through the same gateway and tokens.
Multi-tenant by design
Users, roles, and projects. Agents, knowledge and models live in a project; invite people to it and everyone sees only the projects they belong to.
Tokens & access control
Issue scoped access tokens for chatbots and gateway traffic, pinned to the models they may use. Rotate, revoke, and audit from the dashboard.
Edge first, servers where needed
The API and dashboard run on Cloudflare's edge, with D1, Vectorize and R2; agents that need a filesystem run on our own servers. Anonymous sessions are cleaned up after idle days.
What's built on it
Xynocast builds its business tools on this layer, so each product gets the agents, knowledge, models and voice without rebuilding them.
Trelvia AI
A conversational layer over the ERP and CRM a business already uses — CRM, accounting and customer-support agents. Its models run through this gateway.
Sitehub
Websites their owners change by chatting. The agent harness does the editing; the platform keeps the conversations, the media and the access.