Rockot
One platform for agents, workflows, and enterprise knowledge.
Token Theory designed and built Rockot — a production-grade AI agent platform that consolidates chat, visual workflow orchestration, RAG-powered knowledge bases, and enterprise integrations into a single, extensible system. Multi-model, multi-tool, built to ship fast.
Rockot
Enterprise AI Platform
4 weeks
2025
The opportunity
Organisations adopting AI face a fragmented tooling landscape. Chat interfaces live in one product, workflow automation in another, knowledge bases in a third, and enterprise integrations require custom glue code for each system. The result: teams switch between five tools to accomplish what should be a single workflow, and critical context gets lost at every handoff.
Rockot was designed to collapse that fragmentation into one platform — where an agent can chat with a user, query Snowflake for live data, search an internal knowledge base for policy context, execute a multi-step workflow, and push results to Slack, all within a single conversation thread.
What we built
Agent chat with streaming and tool calling
The core chat interface supports real-time token streaming, multi-model selection (GPT-4o, Claude, Grok), and structured tool calling via the Model Context Protocol (MCP). Agents can reason step-by-step with Claude's extended thinking, analyse uploaded images, and execute code — all within a persistent conversation with full history.
RAG-powered knowledge bases
A complete retrieval-augmented generation pipeline: upload documents (CSV, JSON, PDF, text), chunk them with configurable overlap, generate vector embeddings via OpenAI, and store them in PostgreSQL with pgvector. Background workers process ingestion asynchronously via BullMQ, and semantic search injects relevant context directly into agent prompts.
Visual workflow orchestration
A custom Langflow integration provides a no-code visual builder for multi-step, multi-agent workflows. Connect tools and logic visually, execute flows via REST API, and track metrics (token usage, cost, latency) per run. Scheduled execution via cron lets workflows run daily, weekly, or on custom intervals with full audit trails.
- Slack — send messages, list channels, manage direct messages
- Snowflake — execute SQL queries, explore schemas, export data
- Databricks — SQL execution, cluster management, job monitoring
- Braze — campaign management and customer engagement
- Tavily — real-time web search via MCP
Analytics and execution tracking
Every agent run — chat, workflow, or scheduled — is tracked with token counts, costs, latency, and success rates. A KPI dashboard surfaces trends over time, and TanStack Table provides server-side filtering across the full execution history.
Built with
Architecture
Tech Stack
The impact
Rockot gives organisations a single pane of glass for AI agent operations — replacing a patchwork of disconnected tools with one coherent platform that handles chat, knowledge, workflows, and enterprise integrations natively.
- Six enterprise integrations shipped in 10 weeks (Slack, Snowflake, Databricks, Braze, Tavily, MCP)
- Multi-model support across OpenAI, Anthropic, and xAI with encrypted credential storage
- Full RAG pipeline with pgvector — no separate vector database required
- Background job processing via BullMQ for non-blocking knowledge ingestion and scheduled workflows
- Production-deployed on Vercel with Docker self-hosted option for on-premises requirements
Related projects
Interested in working together?
Let's discuss what's possible for your organisation.
hello@tokentheory.ai