Build an agent

Create your own agent

Buzz ships with BharatRouter built in — AI models hosted in India, wired straight into the app. Create an agent, pick a model, and it's a teammate in your channels. No API key to hunt down, no external account.

4 min read Buzz guide

Agents are AI teammates that live inside Buzz — already met Fizz, Honey and Pollen from your welcome team? You can build your own: a channel helper, a summariser for a busy stream, a first-draft writer for your team. This guide walks through creating one on the desktop app.

Create the agent

Open Agents

In the desktop app, go to Agents and select + → Create agent. Give it a name and, if you like, a persona and avatar.

Point it at BharatRouter

Under AI configuration → Customize for this agent, keep the harness on Buzz Agent (default) and set LLM provider to BharatRouter. Leave the API key field blank — access comes with your workspace.

Pick a model & start

Choose a model — gpt-oss-120b is the recommended default: fast, hosted in India, and solid across everyday tasks. Select Add agent, then Start it from its card.

Put it to work

  • Add it to a channel — from the channel's member list, or from the agent's ⋮ menu. Then @mention it with a question or task; it replies in the flow, visible to everyone.
  • Or DM it for a private back-and-forth — drafts, quick lookups, thinking out loud.
  • By default only you can instruct it ("Who can send instructions" → Only me). Widen that deliberately when the agent is ready for teammates.

What agents are good at

  • Questions & explanations — right in the conversation, no tab-switching.
  • Summaries — a long channel or thread, boiled down with the decisions pulled out.
  • Drafting — announcements, replies, notes; you edit, it types the first 80%.
  • Structured extraction — turn a messy thread into a table, list or JSON.
  • Hindi & Hinglish — ask in the language your team actually types in.

Be specific and they get sharply better: "Summarise the last 20 messages in #launch-room and list the decisions" beats "what's up?" every time. For deep multi-step reasoning or large coding tasks, expect a helper, not an oracle — pick a bigger model and double-check the output.

Agent replies say "Model unavailable"? Switch the agent to gpt-oss-120b in its settings — or if that doesn't clear it, your workspace's BharatRouter access hasn't reached this device yet; ask in your team channel or ping an admin. You never need to buy or paste a personal API key.

Good to know

  • Agents run from the desktop app today. Create and host them on your Mac or Windows machine; everyone can then talk to them from any device — mobile and web included — wherever the agent's been added. (Server-hosted agents that run without your desktop are on the roadmap.) Want an agent with its own name inside Claude Code or Codex instead? See Route B in the setup guide.
  • An agent stops when your desktop app quits, and idle agents pause themselves. Start them again from the Agents screen.
  • Usage is governed by the workspace — models, access and spend are managed centrally. No personal billing, ever.

Want the day-to-day usage patterns — asking well, handing off tasks? That's Working with agents.