# Now (/docs/about/now) --- type: Document title: Now description: The technology Raza works with today, and why each piece is in the stack. Projects and clients are deliberately absent. status: stable order: 4 generated: { by: "process:claude-code", at: 2026-08-27T09:00:00Z } verified: - { by: "human:rrizwan98", at: 2026-08-28T14:00:00Z } ksor: audience: [public] owner: human:rrizwan98 approval: { by: "human:rrizwan98", at: 2026-08-27T09:30:00Z } trust_tier: human-reviewed build_id: sha256:d9f989bee8fa4c15cf0f7f1572f9e9a18305ef4e2ce0c3cfb6a5a421ead6dc1d dirty: true --- What Raza is working with right now, and why. This page names technology, never projects, clients or the systems built on top of it. It is the page that goes out of date fastest, so it carries its own date. **As of 27 August 2026.** ## The runtime for AI employees: OpenClaw An open runtime where each AI employee has its own workspace, its own written instructions, its own skills, its own tools through MCP, and its own channels such as WhatsApp or Slack, and where work can run on a schedule as well as on demand. Why it is the current choice: building an AI-native company needs staff, not chat windows. A runtime that treats the employee as the unit of work gives each one a durable place to keep its instructions and its state, lets several employees hand work to each other, and leaves a place to put the human approval gate. A row of chat sessions gives none of that. ## The way processes reach any AI: MCP Model Context Protocol is the open standard that lets ChatGPT, Claude, Codex and Claude Code attach to an external server. It is how a company's process reaches whichever AI a team already uses, instead of being re-typed into a chat. Servers are written in Python with FastMCP, deployed as remote HTTP endpoints, and hold their state in Postgres so a session that resets can resume. ## The clients on the other side Claude Code, Codex CLI, the ChatGPT desktop app and Claude. The work is deliberately client-neutral: the same connector has to behave the same way whichever one a team prefers. ## Agents and services - **OpenAI Agents SDK** for agents that decide the next step and call tools - **FastAPI** for the services those agents and connectors sit behind - **Docker** for anything that has to run the same on any host ## Data and retrieval - **Neon Postgres** as the store, with **pgvector** where meaning-based search is needed rather than keyword search - **Gemini embeddings** (`gemini-embedding-001`, 1536 dimensions) to turn text and questions into vectors for that search ## Publishing and hosting - **Vercel** for both static sites and container services behind one domain - **Next.js** where a site needs to be more than static pages ## Written procedure as a first-class artifact Skills: written procedures kept verbatim and versioned, read by an agent before it acts. The same procedure runs in every runtime, so two agents cannot drift apart on how a job is done. ## Governed records The record you are reading is itself the current practice: knowledge kept as governed markdown with its sources and approvals, published as a site for people and as an MCP surface for agents, so both answer from the same words. Answers carry citations, and a record that does not cover a question is supposed to say so. ## What this page will not say Which employees, connectors or skills exist in production, how many there are, which clients they serve, or what is inside them. That boundary is described on [what I build](what-i-build.md), and the live connectors are named, with what each refuses, in the [2026 chapter of the journey](../journey/2026-ai-employees-and-connectors.md).