Architecture
Urai breaks a single user prompt across a small team of specialised agents. A Decider routes intent. A Coding Agent writes a tiny program. The Execution Engine runs it in a sandbox. A Summariser turns the result back into a usable answer.
Classifies the request: simple chat, knowledge lookup, or a task that needs code.
Generates a small program in TypeScript using your available tools and knowledge.
Runs that program in a sandboxed serverless environment with a 2-second cold boot.
Reads execution output and produces the response the user sees, with citations.
Splitting work across specialised agents and running compiled code instead of long token chains is what produces our headline numbers: 3x faster, 10x cheaper than a single reasoning agent on the same task, with deterministic outputs you can audit.
UraiJS Runtime
The @tool decorator turns any TypeScript method into an AI-callable function. No SDK boilerplate. No config files. Standard TypeScript with ESM imports. The runtime handles the rest.
@tool One decorator. Any class method becomes callable by AI agents.
meta.secrets Built-in secrets management. API keys encrypted at rest with AES-256.
TypeScript Standard ESM imports, types, async/await. Use any npm package.
Secure Code Execution
When the coding agent writes a program, it has to run somewhere. Urai runs it in a fresh sandbox per task: disposable, networked only to the tools and data you've allowed, gone the moment the task completes.
No persistent state between tasks. Each run is isolated, then torn down. Side effects can't leak across tenants or users.
Fast enough to use inside an interactive chat or a live phone call without making the user wait.
Explicit allowlists for network endpoints. No arbitrary disk access. Auditable I/O for every task.
Coding Agent
Not every team has AI engineers. Urai's built-in coding agent generates UraiJS tools from plain-language descriptions. Describe a workflow, get deployable code.
The agent understands your integrations, your data schema, and your existing tools. Teams go from idea to deployed agent without writing boilerplate.
I need a tool that looks up a customer in Salesforce and creates a Jira ticket with their open issues.
I'll create a tool that connects both systems. Here's the implementation:
Knowledge Hub
Ingest documents in any format. Hybrid search combines semantic understanding with keyword precision, and feeds directly into your agent tools, so agents have context from your documents alongside live data from integrations.
PDF, DOCX, XLSX, PPTX, Markdown, HTML, CSV, JSON, images with OCR, and more.
Semantic vector search meets full-text keyword search. Exact matches with meaning-based retrieval, together.
Fine-grained permissions tied to your identity provider. Users only see what they're authorized to see.
Integrations
100+ pre-built integrations that your agents can use out of the box. Each one becomes tools your agents call: pulling data from Salesforce, creating Jira tickets, posting to Slack.
Channels
Once your agents have tools and knowledge, deploy them through voice or embeddable chat. Same agent, multiple channels.
Natural, interruptible voice agents that take action. Integrate with Twilio, FreeSWITCH, or WebRTC. Sub-second latency.
Embeddable widget for customer chat. Team workspace for internal chat. Project-scoped, model-agnostic, with full access to your knowledge hub.
Agents execute in a sandboxed environment. No arbitrary code on your infrastructure. Secrets encrypted at rest with AES-256.
Sub-second execution for real-time voice and chat interactions where every millisecond matters.
Bring OpenAI, Anthropic, Gemini, or open-source models. Your tools work with any LLM. No vendor lock-in.