Open source · runs on Herdr · coming soon

Wheems: Your own little cloud of helpers.

Wheems is an open-source framework for running a team of AI agents. Hand your chief of staff a job from Slack. Its crew of agents (Claude Code, Codex and opencode, each live in its own terminal on Herdr) does the work, keeps to the guardrails, and only asks you when it matters.

What it’s made of

Five layers, one install.

You talk to a chief of staff. It hands the work to agents that run unchanged, with skills, rules and a knowledge base scoped to the workspace, all live on Herdr.

Works with

  • Claude Code
  • Codex
  • opencode
  • Slack
  • Telegram
  • Herdr
Wheems’s layers, from the top: you, talking from Slack or Telegram; a chief of staff for each workspace; agents on Claude Code, Codex and opencode; scoped skills, rules and the knowledge base; and Herdr, where every agent runs live in its own terminal.

How it fits together

One request, from Slack to the knowledge base.

Every workspace runs the same way. Watch one request go through it, or pick a step.

The opinions

Four opinions, decided for you.

You can rename or tune each one, but not remove it. That’s what makes it a framework.

  1. A chief of staff in every workspace

    Each business or area of your life is its own workspace, with its own agents, skills and knowledge. You talk to one chief of staff per workspace, from Slack or Telegram. It knows every project there, and nothing outside it.

  2. The knowledge loop

    Every session ends with a note. Notes land in an inbox and are filed into the knowledge base for you. Agents never edit it directly, so it stays something you can trust.

  3. Scoped skills

    Skills live in a library that’s never loaded into an agent. Each workspace picks and pins its own set, and each agent picks from that. Nothing leaks everywhere, and nothing changes under you.

  4. Guardrails outside the agent

    Every agent has a shell, so asking it nicely enforces nothing. Rules are written once and enforced where the agent can’t switch them off: GitHub, git hooks, the broker and each workspace’s permissions.

Not another agent loop

Herdr is the engine. Wheems is the missing layer on top.

Chatting with one Claude Code session from Slack is useful on its own. Wheems is the set-up around it: a workspace for each part of your life, a chief of staff in each, agents from different providers, a knowledge base that keeps itself up to date, and rules the agents can’t switch off.

AspectTypical agent harnessesWheems
The agentTheir own agent loop calling a modelClaude Code, Codex and opencode, unchanged, each in its own terminal
ImprovementsArrive with the harness’s releasesEvery Claude Code or Codex update reaches every agent
Seeing the workLogs and chat repliesEvery agent is live in Herdr, where you can step in
ScopeOne instance per businessOne install, many isolated workspaces
Set-upMany choices left openOne way to do each thing
Cross-checkingOne providerAgents from different providers check each other

Getting started

From install to your first chat in Slack.

Install it once, connect Slack, create a workspace. Its chief of staff answers from your phone, and every conversation is a Claude Code session you can watch or take over.

Preview
$ npm install -g wheems
$ wheems setup
✓ Created ~/wheems
✓ Slack connected
$ wheems init acme
✓ Workspace acme, kept apart from the others
✓ Chief of staff on Claude Code, live in Herdr
$ wheems doctor
✓ Node.js  ✓ Herdr core  ✓ Claude Code  ✓ Slack  ✓ Service
Preview. Wheems isn't released yet, and these commands may change before it is.

Helpers included.

Where it is, November 2026: milestone 1 of 9, chatting with a chief of staff from Slack, is built and in daily use. The code goes public under Apache 2.0 when it’s ready for others. Until then, the build log says what got built each week, and why.

Or follow @mikejdeeming on X