harnsy and other tools

This page is written by the makers of harnsy.

People ask how harnsy differs from the tools they already know. This page answers by job, names the tools plainly, and says where each one fits. Every fact about another tool has a date and a link. If something is wrong, write to hello@harnsy.dev and we fix it.

harnsy is a developer tool that connects the Claude Code, Codex and OpenCode sessions you run into a team, and asks you when it needs you.

Which tool for which job

Jobs overlap, and plenty of people use two of these together.

What each tool says it is

Checked on the date shown, on the page linked.

Conductor. “Run a team of coding agents in the cloud.” Each agent gets a cloud sandbox (an isolated microVM with your repos), workspaces can be shared with teammates, and it runs the Claude Code, Codex, Cursor and OpenCode agents on your own subscriptions and keys. Read 31 Oct 2026: conductor.build.

Superset. “One workspace for Claude Code, Codex, and any coding agent.” A desktop app for macOS (experimental Linux, no Windows yet); each agent works in its own git worktree. Source-available under the Elastic License 2.0. Read 31 Oct 2026: superset.sh.

cmux. “The terminal built for multitasking, organization, and programmability.” A macOS terminal. The app is GPL-3.0-or-later; its server parts are under the Business Source License 1.1. Read 31 Oct 2026: cmux.com and github.com/manaflow-ai/cmux.

Claude Squad. Manages several terminal agents (Claude Code, Codex, OpenCode, Amp, Gemini and others such as Aider) in separate workspaces from the terminal, each in its own git worktree. AGPL-3.0; needs tmux and gh. Read 31 Oct 2026: github.com/smtg-ai/claude-squad.

Hermes Agent (Nous Research). “The self-improving AI agent”: its own memory and learned skills, a gateway to chat apps (Telegram, Discord, Slack, WhatsApp, Signal and others), scheduled jobs and sub-agents. Its docs describe a bundled skill that runs Claude Code as a helper on a task, with related skills for Codex and OpenCode. MIT. About 259k GitHub stars on 31 Oct 2026. Its README lists these places to run its commands: local, Docker, SSH, Modal, Daytona and Vercel Sandbox, plus plugin backends such as Singularity/Apptainer. Read 31 Oct 2026: github.com/NousResearch/hermes-agent and hermes-agent.nousresearch.com/docs/user-guide/skills/bundled/autonomous-ai-agents/autonomous-ai-agents-claude-code.

OpenClaw. “an open-source AI assistant that runs on your own computer and meets you in the channels you already use”: Discord, iMessage, Slack, Teams, Telegram, WhatsApp and 20+ more, with apps for macOS, iOS, Android, Windows and Linux. Its README says models and agent harnesses (Claude, Codex, local models) are plugins. MIT; stewarded by the OpenClaw Foundation. About 394k GitHub stars on 31 Oct 2026. Read 31 Oct 2026: github.com/openclaw/openclaw.

Claude Code agent teams. “Coordinate multiple Claude Code instances working together as a team, with shared tasks, inter-agent messaging, and centralized management.” One Claude Code session leads, and its teammates are Claude Code sessions. Its docs call the feature experimental and disabled by default, and document separate cross-session messaging between the Claude Code sessions you run yourself. Read 31 Oct 2026: code.claude.com/docs/en/agent-teams.

Gas Town. “Multi-agent orchestration system for Claude Code, GitHub Copilot, and other AI agents with persistent work tracking”: a coordinator called the Mayor, worker agents, and work state kept in hooks and a ledger called Beads. Its README names Claude Code, GitHub Copilot, Codex and Gemini. MIT. About 18.6k GitHub stars on 31 Oct 2026. Read 31 Oct 2026: github.com/steveyegge/gastown.

halv.ai. “Open source. Half the spend.”: one desktop workspace for Claude Code, Codex, Kimi and GLM on the plans you already pay for, that aims to lower model cost. The app and its local savings engine (output filtering, context compression, a code index) are open source under Apache 2.0, for macOS, Windows and Linux. A paid part, Halv Cloud ($10 a month on its page), adds routing that sends each task to a model, a cloud savings dashboard and account rotation. It publishes a benchmark of 42 paired runs with its data. Read 31 Oct 2026: halv.ai.

OpenRig. A terminal tool on tmux: you define an agent team in YAML and start it with one command; Claude Code, Codex and Pi today. macOS and Linux (Windows via WSL2). Apache 2.0. Read 31 Oct 2026: github.com/mvschwarz/openrig.

Where other tools fit

• Open source code. Hermes Agent, OpenClaw and OpenRig publish their code (MIT, MIT, Apache 2.0), and halv.ai its app and local engine (Apache 2.0; its cloud part is not open); Claude Squad (AGPL-3.0) and the cmux app (GPL-3.0-or-later) do too. harnsy's source code is not published; the installers are public.
• Community size. Hermes Agent has about 259k GitHub stars and OpenClaw about 394k (31 Oct 2026). harnsy is new and small.
• An assistant with its own memory and skills. Hermes Agent describes its own memory and learned skills. harnsy leaves each tool's own memory (CLAUDE.md, AGENTS.md) as it is and adds project memory on top: roles, project documents, decisions on tasks, handover notes and task history.
• Many chat apps. OpenClaw lists 20 or more channels; Hermes Agent lists Telegram, Discord, Slack, WhatsApp, Signal and others. harnsy has Telegram and Discord with harnsy Max.
• Places to run commands. Hermes Agent’s README lists Docker, SSH, Daytona, Modal and other places to run its commands. harnsy runs agents on your own machines (with harnsy Max, on any machine of your cluster) and has no disposable per-task sandbox; a task can run in its own git worktree, which is not isolation.
• Lower model cost. halv.ai aims to lower model cost with output filtering, context compression and a code index on your machine, and task routing in its paid Halv Cloud. harnsy does not filter output, compress context or route tasks by cost.
• Built into Claude Code. Claude Code’s own agent teams and cross-session messaging work inside Claude Code (its docs call agent teams experimental and disabled by default).
• Simple parallel workspaces. If you need several agents side by side, the parallel-workspace tools are the simpler choice.

What harnsy is built for

• Claude Code and Codex sessions you start yourself join the team; OpenCode sessions are started through harnsy.
• Claude Code, Codex and OpenCode in one conversation: a message lands in the running session.
• Questions, permission requests and acceptance on a task card wait for you. With harnsy Max you also get phone push, Telegram and Discord (see “Where does my data go?” in the FAQ for what those services see).
• Roles, project documents and handover notes carry a seat over to a new session after a restart or a full context window. Since 0.10.0 each post also keeps lessons approved by the lead or you; the next holder reads them when it starts.
• OpenCode agents can run other models next to Claude Code and Codex; harnsy Max adds provider connections.

Using them together with harnsy

What we have checked:

• A Claude Code or Codex session you started yourself (in your terminal or in tmux): install harnsy and it finds the session, shows it and delivers messages into it. To make it a team member, tell it to read the harnsy guide and give it a seat. OpenCode sessions need to be started through harnsy.
• OpenClaw, Hermes Agent: harnsy has no connection to them. Keep them for what they do and run the coding team in harnsy next to them.
• GitLab or GitHub issues: harnsy does not import or sync them. Keep your tracker; the harnsy board holds the agents’ own work items, with your acceptance.
• Kimi and GLM: in harnsy they run as OpenCode agents; harnsy Max adds provider connections.

Questions we get

1. Is harnsy the same as Conductor, Superset, cmux or Claude Squad?

They overlap on one point: all of them let you run several coding agents at once. Those tools give each agent its own workspace (Superset and Claude Squad, for example, a git worktree per agent; Conductor, a cloud sandbox). harnsy has a different job: it connects the sessions you run into one team that messages each other, hands work over and asks you when it needs you. If you need agents side by side, the workspace tools are the simpler choice. Many people use both.

2. Hermes Agent already runs Claude Code and Codex. Why harnsy?

True overlap: Hermes Agent’s docs describe a bundled skill that runs Claude Code as a helper on a task, with related skills for Codex and OpenCode. The job is different. Hermes Agent is an assistant with its own memory, learned skills and chat apps. harnsy runs a coding team of Claude Code, Codex and OpenCode sessions, with you asked when needed. harnsy has no connection to Hermes Agent today, so keep it for what it does and run the coding team in harnsy next to it.

3. Is harnsy like OpenClaw?

Both can talk to you in chat apps, and both can work with Claude and Codex. The job differs: OpenClaw describes itself as an open-source assistant that runs on your own computer and meets you in the channels you already use. harnsy runs a coding team. If you want an assistant across many chat apps, OpenClaw lists 20 or more channels; harnsy has Telegram and Discord with harnsy Max. harnsy has no connection to OpenClaw, so people who want both run both.

4. OpenRig is open source. Why harnsy?

OpenRig is a close neighbour: it also connects Claude Code and Codex into a team. It works in the terminal, with the team defined in YAML, and its code is published under Apache 2.0. harnsy has a web dashboard, a task board whose results you accept, and with harnsy Max a phone app and Telegram and Discord; its source code is not published. If you want the source, or you live in the terminal, OpenRig fits. Using both is possible.

5. halv.ai already puts Claude Code and Codex in one app. What is different?

Both put Claude Code and Codex in front of you, and halv.ai is one desktop app for them. Its job, as it describes it, is to lower model cost with output filtering, context compression and a code index on your machine, and task routing in its paid Halv Cloud. harnsy does not do those things; its job is to connect the sessions you run into a team. If cost is what you want to cut, halv.ai is built for that. You can run both.

6. Claude Code already has sub-agents. Why another layer?

Claude Code’s agent teams do overlap with harnsy: one Claude Code session leads and its teammates are Claude Code sessions, and its docs describe messaging between sessions you run yourself (the docs call agent teams experimental and disabled by default). harnsy connects Claude Code, Codex and OpenCode sessions you have already opened into one team, with roles and a board. If Claude Code is the one tool you use, the built-in features may be all you need. You can use both.

7. Why is harnsy’s source code not published?

It is not: harnsy is closed source under “All rights reserved”, and the installers are public. That is a real difference from Hermes Agent, OpenClaw and OpenRig, which publish their code, and halv.ai, which publishes its app and local engine, and from Claude Squad and the cmux app. If you need the source, those are the tools to use.

8. Do I have to pay?

harnsy runs without a key: 2 projects in one organization, with any number of agents. harnsy Max features are turned on by a key.

9. Can I bring my setup (Hermes Agent, OpenClaw, GitLab)?

Partly. Claude Code and Codex sessions you started yourself in your terminal or tmux are found by harnsy and can join the team. harnsy has no connection to Hermes Agent or OpenClaw, so keep them for what they do and run the coding team next to them. harnsy does not import or sync GitLab or GitHub issues: keep your tracker, and the harnsy board holds the agents’ own work items, with your acceptance.

10. I need more chat apps, or a sandbox per task.

Hermes Agent and OpenClaw list more chat apps than harnsy, which has Telegram and Discord with harnsy Max (see “Where does my data go?” in the FAQ for what those services see). A per-task sandbox is not in harnsy: it runs agents on your own machines, and a task can run in its own git worktree, which is not isolation. Hermes Agent’s README lists Docker, SSH, Daytona, Modal and other places to run its commands. Use them next to harnsy if you need that.

Last full check

Last full check: 31 Oct 2026. Facts about other tools were read on their own pages on the dates shown on each card. Facts about harnsy were confirmed against harnsy 0.10.0 (released 19 Oct 2026) on 31 Oct 2026. Found a mistake? hello@harnsy.dev.

Names are used only to identify the tools. harnsy is not affiliated with, or endorsed by, any of them. All names belong to their owners.

---