A multiplayer chat agent for your team.
Most AI assistants are single-player. heypi puts one agent in your team's Slack, Discord, or Telegram, with sandboxed tools, approvals, and an audit trail for anything sensitive. An app you own, not a workflow platform.
Your agent is just files
No dashboard to configure. An agent is a folder you edit and commit to git. The scaffolder writes it for you.
- instructions.md: who the agent is and how it behaves.
- tools/: TypeScript functions it can call.
- skills/: runbooks it can follow on demand.
- schedules/: scheduled and recurring work.
- index.ts: chat platforms, runtime, and state.
Everything a shared agent needs
heypi wraps a Pi agent with what a team actually needs: a place in chat, governed tools, durable state, and a record of everything it does.
Chat platforms
Add Slack, Discord, Telegram, or a trusted webhook. One agent, shared by everyone.
Approvals
Gate sensitive actions behind human approval, with named approvers for your team.
Audit trail
Inspect durable conversation events, approvals, schedule runs, and Pi session transcripts.
Files & commands
Let the agent read and write files, run commands, and search a scoped workspace.
Sandboxing
Use host or Docker execution, or plug in Gondolin, just-bash, Vercel, Cloudflare, or a custom runtime.
Memory
Store curated durable notes and retrieve them only when they are useful.
Skills
Stage runbooks, scripts, and assets that Pi can load on demand.
Secrets
Collect credentials through an encrypted handoff, never pasted in chat.
Scheduling
Run code-owned cron schedules as isolated prompts or dispatch into a configured conversation.
File attachments
Receive files from chat and send generated files back.
Admin panel
Inspect conversations, approvals, active work, Pi sessions, schedules, and secret requests.
Custom tools
Add trusted TypeScript tools for your own apps, data, and APIs.
FAQ
What is heypi?
heypi is a TypeScript framework for multiplayer chat agents: one Pi agent your whole team uses in Slack, Discord, Telegram, and trusted webhook entrypoints, with approvals, audit trails, scoped runtime tools, memory, secrets, schedules, and an admin UI.
How do I use heypi?
Run npm create heypi@latest -- codex-tag my-agent. It copies the Codex Tag template, installs its dependencies, and gives you a working agent to configure. You can also build manually with loadAgent(), runHeypi(), and an adapter.
Is heypi a workflow platform?
No. heypi is optimized for chat-ops agents with governed tool access. It records interrupted work during startup recovery, but it does not replay arbitrary in-flight agent turns after a process crash.
What can I connect it to?
The core package supports Slack, Discord, Telegram, local, and webhook adapters. See adapters, or write another transport with the custom adapter guide.
Which models can I use?
heypi delegates model calls to Pi. Configure the model in agent configuration and provide provider credentials, such as OPENAI_API_KEY, through the Node process environment.
Do I need Docker or a VM?
No. Host execution is enough for many agents. Use Docker, Gondolin, Vercel, Cloudflare, or a custom runtime provider when the agent needs a real container or remote execution boundary.
Can agents ask before acting?
Yes. Tool calls can require approval before they continue. heypi records who requested the action, who approved it, what ran, and how it ended. See tool approvals.
What does heypi store?
Your heypi app stores its state on your own infrastructure. heypi keeps append-only conversation and audit records, schedule state, workspaces, and encrypted secret handoffs under its configured state directory, while Pi owns its session transcripts. heypi.dev is only the website and docs.
How much does heypi cost?
heypi is free and open source. You still pay for your own infrastructure, model provider, chat platform, and any external runtime provider you choose.
Is heypi open source?
Yes. The code is on GitHub, packages are published on npm, and the license is in the repository.
Who built this?
heypi is built by Ronan Berder.
Why did you build this?
Most AI assistants are single-player: one person, one private chat. Teams need something different. A shared agent in a shared channel needs channel adapters, scoped context, runtime boundaries, approvals, audit logs, and secret handoff. heypi packages those pieces around Pi.
How can I help?
Star the project, report issues, send pull requests, or sponsor development.