GitHub - xatuke/wrapper: wrap your agents

Host, manage, and observe pi agents on Slack — each agent in its own hardened container, with Langfuse tracing (tokens, cost, tool spans), managed by one CLI.

An agent is config + tools; everything else is a shared runtime. Adding an agent means writing an agent.yaml, optionally dropping tool modules in a folder, and running wrapper agent add — not writing a new service.

agents/support-bot/
├── agent.yaml     # model, prompt, channels, tools, features
├── prompt.md      # system prompt
├── tools/*.ts     # custom tools (TypeBox schema + execute)
├── slack/         # slack-cli project (manifest.json, app registry)
└── .env           # Slack tokens + keys (gitignored)

Each agent gets its own Slack app (Socket Mode — outbound only, no public ingress), a persistent volume for conversation history, and per-turn traces. Opt-in runtime features give agents self-managed cron schedules, persistent memory, and a file workspace. Custom Dockerfiles bake in extra CLIs or npm packages; multi-line secrets go in base64 via wrapper env set-file.

Quick start

npm install --ignore-scripts && npm run build
cp wrapper.example.yaml wrapper.yaml          # set your Slack team ID
alias wrapper="node $PWD/packages/cli/dist/index.js"

wrapper init --team-id T0XXXXXXXXX            # starts Langfuse, provisions org
echo 'ANTHROPIC_API_KEY=sk-ant-...' >> agents/.env

wrapper agent add support-bot --channels "#support"
wrapper up support-bot                        # then /invite @support-bot in Slack

Full walkthrough: docs/getting-started.md.

Documentation

Doc	What
docs/getting-started.md	Local setup end to end (prereqs, init, first agent, day-to-day)
docs/agents.md	Creating agents: `agent.yaml` reference, custom tools, cron/memory/workspace features, custom images
docs/langfuse.md	Tracing: the bundled stack, using an existing Langfuse deployment, or Langfuse Cloud
docs/deploy-vm.md	Running 24/7 on a plain cloud VM (AWS EC2 / GCP GCE) — no Kubernetes
docs/deploy-k8s-aws.md	Kubernetes on AWS: EKS, ECR, Bedrock via IRSA, hardening
docs/deploy-k8s-gcp.md	Kubernetes on GCP: GKE, Artifact Registry, Workload Identity, hardening
docs/k8s-manifests.md	The generated Kubernetes manifests, annotated, and how to customize them
DESIGN.md	Architecture and dated design decisions
AGENTS.md	Orientation for AI coding agents working in this repo

Deploy targets

wrapper.yaml sets target: local | k8s — same commands either way:

local — docker compose on your machine or a cloud VM, with a bundled self-hosted Langfuse (or your existing one).
k8s — wrapper up builds and pushes images (ECR or Artifact Registry) and deploys each agent as a hardened Deployment (no Service/Ingress) with a PVC for sessions and a Secret synced from the agent's .env.

Day-to-day

wrapper ui                      # web console at localhost:3020 (status, logs, env)
wrapper status                  # containers/pods + langfuse health
wrapper logs support-bot -f
wrapper env set support-bot K=V && wrapper restart support-bot
wrapper agent remove support-bot [--purge]

What's tracked where

This repo is the framework. docker-compose.yml, k8s/, and each agent's .slack/ app registry are generated — gitignored, regenerated by the CLI, never edited by hand. wrapper.yaml (deployment config) and wrapper.secrets.json are yours and gitignored too; create them from wrapper.example.yaml / wrapper.secrets.example.json.
Your agents are yours. The entire agents/ folder is ignored by wrapper's git — keep your agents (and optionally your wrapper.yaml) in their own repo nested inside it, so git pull here updates the framework without touching them. wrapper agent add scaffolds the folder with a README and secret-excluding .gitignores; see docs/agents.md.

A fresh machine needs: clone wrapper, clone your agents into agents/, npm install --ignore-scripts && npm run build, a wrapper.yaml, wrapper init, restore .env secrets, wrapper up.