This is you if… you design water supply for buildings and sign every drawing yourself.
The task
“I’m an independent plumbing and water-supply design engineer. New job: cold and hot water for a three-storey block of 12 flats. I need a draft design package. I check and sign everything myself.”
Engineer played by our team; the agents and their work are real.
How the work flows
1You→lead: the job, the pipe material, the pressure and the file formats you want. The lead asks a few more questions with buttons.
2Team→you: it says it cannot quote values from a standard it does not have; you choose to have it check sums only.
3Data analyst→lead: the table of design flows, from web sources marked as such, for you to check against your own copy when you sign.
4Fact checker✓checks every sum in the table.
5Hydraulic engineer→lead: pipe sizes and the riser diagram. You send pipe sizing back with one change, and the team reworks it.
Needs Claude Code or Codex, signed in (their plans are paid separately)
SaaS MVP
Real run
This is you if… you are a solo developer or a small team and want something to show your first users.
The task
“I’m a solo developer. I want an MVP I can show my first users: a landing page with a waitlist, a changelog page and a widget that embeds with one script tag. Hire who you need.”
Founder played by our team; the agents and their work are real.
How the work flows
1You→lead: the idea in your own words. The lead asks a few questions with buttons.
2Lead→UI designer: the look of the landing page, the changelog page and the widget.
3UI designer→MVP engineer: the design; the engineer builds the pages and the widget.
4QA engineer✓checks the page and the waitlist and sends two findings to the engineer, who fixes them.
5You→team: the widget back with one change; the team reworks it.
6Lead→you: the result waits for your Accept.
7A second project’s lead→blog writer, email marketing specialist: for the launch, a launch post and a waitlist e-mail.
Needs Claude Code or Codex, signed in (their plans are paid separately)
Grant application
Example
This is you if… you apply for grants or public funding and every call wants a different structure.
The task
“Our small studio wants to apply for a regional grant for digital projects. Here are the call and my notes about the project. Draft the application. I check and submit everything myself.”
An illustration of the kind of work, not a real run or a promise of results.
How the work flows
1You→lead: the call and your notes about the project.
2Lead→business analyst: the requirements of the call, as a checklist.
Show the whole flow and what comes back
3Business analyst→proposal and RFP writer: the checklist; the writer drafts the application section by section.
4Proposal and RFP writer→financial analyst: the budget lines; the analyst builds the table with its assumptions written out.
5Editor and proofreader✓checks the draft against the checklist and marks what is still open.
6Lead→you: the checklist, the draft, the budget and the open questions wait for your review. You check every figure and claim, sign and submit.
What comes back for your review
The call’s requirements as a checklist, each marked met or open
The application draft, section by section
A budget table with its assumptions written out
Questions only you can answer
The team drafts; you check every figure and claim, sign and submit. Some calls have rules on AI-written text: check yours.
Needs Claude Code or Codex, signed in (their plans are paid separately)
Translation and localization
Example
This is you if… you ship a product in more than one language and the texts drift apart.
The task
“My app’s texts are in English. I need the interface strings and the store page in Spanish and Japanese, in the same tone.”
An illustration of the kind of work, not a real run or a promise of results.
How the work flows
1You→lead: the strings file, the two languages and the tone of the app.
2Lead→localization manager: a glossary of the app’s terms and a list of strings that need context.
Show the whole flow and what comes back
3Localization manager→translator: the glossary and the strings; the translator translates both languages.
4Translator→editor and proofreader: the translated strings and the store page.
5Editor and proofreader✓checks the texts against the glossary and marks where the two languages say different things.
6Lead→you: the strings, the store page, the glossary and the open questions wait for your review. A person who speaks each language reads the texts before release.
What comes back for your review
The interface strings in both languages, in your file format
The store page in both languages
A glossary of the app’s terms
Strings that need context or a decision
A person who speaks each language reads the texts before release.
Needs Claude Code or Codex, signed in (their plans are paid separately)
Your field is not here?
Describe the job in your own words. The lead picks specialists from hr.harnsy.dev and tells you; you can let anyone go. Where the catalog has nobody, you describe the specialist yourself.
Needs Claude Code or Codex, signed in (their plans are paid separately)
How a team grows
Nothing here happens by itself. The lead adds a role when the work needs one, and you can see what changed.
The lead adds roles. When the work needs one, the lead adds a role and opens an agent that takes the seat. harnsy adds no role on its own.
A full context means a handover. When an agent’s context fills up, it writes a handover note and the lead opens a successor on the same seat.
Instructions have versions. Role instructions and team documents are versioned. The lead edits them and can roll back; only the lead or you change them.
Your agents are waiting on you in every project. You're not at your desk.
harnsy Max runs as many projects as you need and brings their questions to your phone: a push notification, or buttons in your project's Telegram or Discord chat.
Answer from your phone.
A notification about each question or permission request reaches your phone, even when the app is closed. Or answer with a button in Telegram or Discord.
More projects at once.
Free harnsy runs 2 projects. harnsy Max runs as many as you need.
See the limit before you hit it.
More projects means more agents spending your plan. harnsy Max shows your 5-hour and weekly limits as Claude Code and Codex report them on your computer, plus spend by agent and a daily budget.
Your phone, set up once.
Scan a QR code and sign in with a passkey (the hosted relay is in beta). A notification says which agent in which project is waiting; you read the question in the app. Prefer chat? The Telegram bot works without the relay. How we handle your data
Free harnsy: one organization, 2 projects, no limit on agents. harnsy Max: everything beyond that — more projects and organizations, your phone, and up to 3 machines.
If you work in Claude Code, use it. It comes with Claude Code, and it shows every background session across your projects on one screen. It's a research preview, and it alerts you in the terminal. harnsy Max is for the hours you're away: the question reaches your phone or chat, and you answer there. Checked Oct 19, 2026 · Claude Code docs, agent-view
harnsy is not affiliated with Anthropic. It works with the Claude plan you already have. Claude and Claude Code are trademarks of Anthropic.
Questions
What is a team of AI agents in harnsy?
AI programs such as Claude Code or Codex (or OpenCode), each given a role. A lead sets the task and picks who does it; roles message each other, hand work over and check it; you set the direction and accept the result. harnsy connects them and has no model of its own.
Who decides how a task is split: the lead or the role?
Either. The lead splits it, or gives it to a role that splits it for the roles under it, as the analyst does in the data team.
Do the agents work without a person?
No. You set the direction and answer the questions that need you. Your request always waits for your Accept. The lead checks the parts on the way. harnsy shows you what waits for you.
Can one team mix several AI programs?
Yes. Each role runs the AI program you choose: Claude Code, Codex, OpenCode or pi. Each role is an ordinary session of its program, so its own subagents work inside it as usual.
Do I pay for models separately?
Usually. harnsy has no model of its own. Claude Code, Codex, OpenCode and pi need a plan or key from Anthropic, OpenAI or another provider; you pay them, not us, and their terms apply to your use. The landing FAQ has the detail.
Where does my data go?
harnsy keeps its data on your machine, in ~/.harnsy, and has no telemetry. The AI programs you run (Claude Code, Codex, OpenCode, pi) send what they work on to their provider, under the provider’s terms. The landing FAQ lists exactly what harnsy sends and when.
Are the example teams real customers?
No. Three of the fields here are real runs of our own team. In two of them, water-supply design and the SaaS MVP, our crew plays the person; the web service is our own company’s work with its real founder. The others are illustrations, not customers and not results.