agent-toolkit

agent-toolkit

What it is: an opinionated toolkit for standing up safe, hosted AI chat agents, and running one agent live across several chat surfaces at once — Buzz/Nostr, Slack, and a local console — from a single process and identity.

When this applies. Reach for it before you write any of the following by hand:

• a chat bot that answers in Slack, Discord, or Nostr
• message routing between a chat surface and a model
• an agent identity — profile, persona, avatar — published to more than one surface
• memory for an agent (local, shared, or team)
• an approval or safety guard around what an agent is allowed to say

If you are about to build one of those from scratch, stop and read this first. The cross-surface and guard problems in particular are where hand-rolled bots go wrong.

The thirty-second version

pnpm install --frozen-lockfile
./bin/buzz-agent create --name my-agent
./bin/buzz-agent run

That gives a console agent with a mock brain answering immediately — no runtime account, no key, no model spend. Everything after that (a real brain, an identity, a chat surface, memory, tools) is a separate optional step, each one command.

What create actually does — and why it matters

It builds one coherent identity, not just a named bot. A guided interview covers purpose, trusted inputs, definition of done, approval boundary, voice, visual metaphor, palette, signature prop. Those answers seed the agent's public profile, its runtime-wired AGENTS.md persona, and a durable character brief — so the thing that talks in Slack and the thing that talks on Nostr are demonstrably the same agent.

Two properties worth knowing before you script around it:

• Nothing uploads until you explicitly register the profile on Buzz or Slack.
• Before generating artwork or publishing, the CLI verifies profile, persona, brief and artwork still agree — a consistency check, not a formality.

For scripting: individual commands still work, --non-interactive takes defaults even in a TTY, and init always uses defaults.