Everything a team of your tools needs.
Roles, direct messages, handover, your checkpoints, any model, several machines and subscriptions: what harnsy adds on top of Claude Code, Codex and OpenCode.
Learn how it works: the course →
Persistent roles
Roles with job titles. Each agent holds a role in your project, with its duties, goals and the team’s rules, and the role stays across sessions, whichever agent holds it.
Roles, goals and tasks live in harnsy, not in one session’s head. Every agent that takes a role starts from its document.
Agent-to-agent communication
Messages land right in the recipient’s session. No inbox polling.
Claude Code, Codex and OpenCode send each other questions and answers. Each message lands right in the session.
Documents for agents
Write the team’s rules once, as documents for a project or one role; every agent in the role finds them and reads them when needed.
Project rules, checklists and references live as documents in Roles, for the whole project or one role. The role document links them.
Handovers
When the context runs low, the agent writes a note and a fresh agent takes the role. The work stays.
Agents change; the work goes on. Instead of compressing a tired context, harnsy brings in a fresh agent and keeps the old one as a consultant.
Human checkpoints
Answer your agents right from the dashboard: questions and permission requests come with buttons, and Home shows what else waits for you.
harnsy calls you when a decision is needed, and you give it right there. The rest of the time, you work.
Task management
The lead turns your request into tasks for each role; agents attach proof, and you accept or return the result.
The lead turns what you ask into tasks on the board, splits them by role and brings back the result with proof. You decide on the task card.
Changes
See what every agent is changing: each worktree’s branch, state and full diff, and a warning when two of them touch one file.
Agents work in their own git worktrees. Changes shows each one: its branch, task and agent, whether it’s in master, and the full diff.
Terminals and agents
Every agent in a tab of its own (tmux if it is installed, otherwise harnsy’s own terminal host; on Windows always its own): watch any of them in the dashboard, step in, and bring the team back after a reboot.
Agents run in tmux if it is installed, otherwise in harnsy’s own terminal host (on Windows, always its own), a tab each. The dashboard opens their terminals, reopens the team after a reboot and lists every agent with its role and context.
Notifications and settings
Home and the tray show what needs you; the lead hears about the rest first. You choose which notices go out, and when.
Home and the tray show what waits for you. The lead agent hears about the rest first. System › Notifications decides what goes out, and when.
Any model
Mix Claude Code, Codex and OpenCode. Hand routine work to local or low-cost models.
Claude Code, Codex and OpenCode work in one team. Each does what it’s best at, and you’re not tied to one vendor or its limits.
Dictation
Speak into any text field of the dashboard. The text lands at your cursor, and you edit it and send it yourself.
Every text field has a microphone. harnsy sends your speech to the recognizer you choose and puts the text at your cursor. You edit it, then send it.
Multi-machine teams
With harnsy Max, agents on different machines work as one team. Production gets its own agent, and only it holds the keys.
Laptop, server, production, staging. Each environment has its own agent and its own access. It’s in every harnsy build; a harnsy Max key turns it on.
Access from outside
With harnsy Max, open your dashboard from a phone or another computer over the internet, behind a passkey. A change asks for the passkey again.
Open the dashboard from a phone or another computer over the internet, behind a passkey. It’s in every harnsy build; a harnsy Max key turns it on.
Connections
With harnsy Max, one machine holds several Claude, Codex and OpenCode logins and API keys, and each project and agent runs on the one you pick.
Your Claude, Codex and OpenCode logins and API keys live in Connections, each project and each agent runs on the one you pick, and you switch when a limit runs out. It’s in every harnsy build; a harnsy Max key turns it on.
Usage
With harnsy Max, see every 5-hour and weekly limit with a forecast to the reset, its history, and what your agents spent.
Usage shows every 5-hour and weekly limit of your logins with a forecast to the reset, the history of each limit, and what your agents spent. It’s in every harnsy build; a harnsy Max key turns it on.
Schedules
Give a role a rule, like every weekday at 9. At that time harnsy puts a task on the board, and starts an agent if nobody is there.
Give a role a rule — every weekday at 9, every Friday, once — and at that time harnsy puts a task on the board for it. If nobody is there to take it, harnsy starts an agent.
Actions
Agents write a script for a job you repeat; you read it, enable it and run it from a button, with the steps, log and result in the dashboard.
An action is a script an agent writes for a job you keep doing by hand: a build, a check, a cleanup. You read it, enable it, and run it from the dashboard with one press.
Messengers
With harnsy Max, a Telegram or Discord bot per project: write to the lead from your phone, answer questions with buttons, send voice, and let agents post to the chat.
Give a project its own bot. You write to its lead from your phone, and the lead’s questions reach your private chat as buttons, so you press one. A harnsy Max key turns it on.
Keys for agents
With harnsy Max, agents call GitHub or a model provider through harnsy: the key is added on the way out, only to the hosts you set, and the agent never sees it.
An agent that needs GitHub or a model provider asks harnsy to make the call. harnsy adds the key on the way out and sends it only to the hosts you listed. The key never enters the agent’s session. A harnsy Max key turns it on.
The Assistant
A helper built into harnsy: ask it, hand it a setup job or an idea for a backlog. It proposes, and nothing changes until you press Apply.
The Assistant is built into harnsy. Ask it how something works, to set something up, or to put an idea in a project’s backlog. When a change touches a project’s files, it shows you the change and nothing happens until you press Apply.
Files
One folder tree for the files agents hand to each other and for copies of task attachments, with a viewer, limits and a trash. With harnsy Max: S3 shared by all machines and links with an end date.
The files agents hand to each other, copies of your task attachments and your documents live in one folder tree that you can open in the dashboard. Keep it on disk, or with harnsy Max in an S3 bucket that all your machines share.
Phone app and push
With harnsy Max, install the dashboard on your phone, get a push when an agent waits for you, and use a terminal that fits the screen.
With harnsy Max, install the dashboard on your phone as an app. A question or a permission request from an agent reaches you as a push notification, and you answer from the phone, even with the app closed. It is a web app you install from the browser, not an app from a store.
Lead on every page
A tab and a button on every page open the lead’s chat. A shortcut opens it, and you can dictate instead of typing.
Whatever you are looking at, the lead is a tab and a button away. Ask, decide or dictate, then go back to what you were doing.
External MCP servers
Connect outside services by key from a catalog of 28, review their tools, grant them per project and role, and read the call log.
Connect an MCP server by key, look at its tools, and grant them per project and role. Every call goes into a log you can read.
Relay
With harnsy Max, a main machine with no public address still has members: they join it through a relay. Nobody opens a port, and the relay only copies bytes.
A laptop, a home computer or a server behind NAT can be the main machine of your cluster. The main machine and each member that joins through the relay connect out to it, and the relay only passes bytes between them. A member behind NAT needs no relay: it joins a reachable main machine directly.
Milestones
A milestone is a release of your project, with a due date, a theme and a release task. Assign tasks to it and see how far the release is.
A milestone is a release of your project: a name, a due date, a theme and the release task that ships it. Tasks you assign to it, with their sub-tasks, are what the release is made of. The project roadmap is made from the same milestones.
Export / Import
Save your company to one file: its projects, teams, roles, documents and goals. Open it on another machine, or hand it to someone else, and set up the same team there.
Export your company to a file, and import it on another machine or for another person. The file carries the company with its projects and teams: roles (the company’s role templates and each project’s own changes to them), documents and goals. It carries no secrets, no agent sessions and no tasks.
Your first team is one prompt away.
Installing takes one sentence to your agent. To run your agents, harnsy drives a terminal: tmux if it is installed, otherwise its own terminal host (on Windows, always its own).
Install harnsy following https://harnsy.dev/llms.txtQuestions
What's a harness?
The program an AI agent runs in: Claude Code, Codex, OpenCode. harnsy doesn't replace them; it connects them.
What is harnsy, physically?
A small service on your machine and a dashboard in your browser.
What's in harnsy, and what's in harnsy Max?
harnsy itself is free: 1 organization, 2 projects, no limit on agents, one machine. Claude Code, Codex and OpenCode need their own plans or keys, paid to their providers. harnsy Max is the same build with a key: more projects and organizations, access from your phone (beta), several machines as one team, several subscriptions and their usage.