Your agents do the work. Zenborg checks it.
Stop babysitting your AI. Zenborg is an add-on app that puts a Manager → Worker → Reviewer loop over the agents you already use, so you come back to checked work instead of half-finished drafts.
Works with Claude Code, Codex, Hermes, OpenCode and any agent app that supports MCP. Runs on your machine, under your own accounts.
Free public beta · No credit card · Key arrives in under 60 seconds
Get your afternoons back from AI babysitting.
Describe the task once. Your agents plan, build in an isolated git worktree and review each other. You inspect the diff and approve, once in a while, not every minute.
Works with Claude Code, Codex, Hermes, OpenCode and any agent app that supports MCP. Runs on your machine, under your own accounts.
Free public beta · No credit card · Key arrives in under 60 seconds
One sentence starts a checked agent loop.
Write one sentence inside the tool you already use. Your agents plan it, build it in a safe copy of your project, check the result, and ping you when the tests pass.
Works with Claude Code, Codex, Hermes, OpenCode and any agent app that supports MCP. Runs on your machine, under your own accounts.
Free public beta · No credit card · Key arrives in under 60 seconds
I stopped clicking "Allow" every thirty seconds.
My story. Your results depend on your project and agents.
Zenborg lets my agents check each other before anything touches my files. I set the goal, they run the loop in an isolated worktree, and I review the finished diff. Free during the public beta.
Works with Claude Code, Codex, Hermes, OpenCode and any agent app that supports MCP. Runs on your machine, under your own accounts.
Free public beta · No credit card · Key arrives in under 60 seconds
Put a reviewer over your AI. Stop clicking "yes".
A Manager plans, a Worker builds, a Reviewer checks the diff and the tests. You walk away after one sentence and approve verified work when you are back.
Works with Claude Code, Codex, Hermes, OpenCode and any agent app that supports MCP. Runs on your machine, under your own accounts.
Free public beta · No credit card · Key arrives in under 60 seconds
Run checked agent loops on several projects at once.
Each loop gets its own isolated git worktree, so three projects can progress in parallel without touching each other. You review each finished diff when you are ready.
Works with Claude Code, Codex, Hermes, OpenCode and any agent app that supports MCP. Runs on your machine, under your own accounts.
Free public beta · No credit card · Key arrives in under 60 seconds
- ✓Plan written before any code
- ✓Work done in an isolated git worktree
- ✓Manager reviews the diff and sends fixes back
- ✓Project tests must pass before "done"
Sound familiar?
You finally have momentum on your project, and then the routine starts again.
- The 30-second leash. You give your AI a task. Thirty seconds later, the screen pauses and chimes: "Can I run this command?" You click Yes. Thirty seconds later, another pause. You click Yes again. Your coffee goes cold while you babysit a machine.
- The fake green checkmark. Ten minutes later it announces: "Done, all tests pass." You open the app to check. The button does nothing. Half the feature is missing.
- The four-window relay. You copy the red error text, alt-tab to another agent window, explain the whole project from scratch, copy its guess back, and pray the two agree.
You started using AI to build things faster. Not to sit beside it clicking "allow" all afternoon.
The fix: put a manager loop over your agents.
In a real team, the person who writes the code is not the only one who checks it. Someone plans, someone builds, someone reviews.
Your AI workflow can work the same way. Right now, one agent plans, builds and grades its own homework, and you babysit the permission prompts. Zenborg fixes that with the Manager Loop. It gives the agents you already use three clear roles:
The Planner plans
The agent you already talk to, such as Claude Code or Codex. You tell it what you want in plain English. It maps the plan and defines what "finished" looks like.
The Manager checks
A second agent, such as Hermes or Claude Code. It inspects the work, catches edge cases, and sends mistakes back for repair before you ever see them.
The Worker builds
For example, Codex or OpenCode. It writes the files inside an isolated git worktree, so your real files stay untouched until you review and approve.
And Zenborg enforces verified completion: a loop never marks itself "done" until your project's actual tests pass.
You type one sentence. You walk away. You come back to finished, verified work.
Your first loop takes 60 seconds. Literally one sentence.
No dashboards to wire up. No config files to edit. When Zenborg opens for the first time, it automatically connects itself to the AI agents already installed on your Mac or Windows PC.
Open the agent you already use (Claude Code, Codex, Hermes, or OpenCode) and paste this single sentence:
Use Zenborg. Plan this task and start a Zenborg loop with [e.g. Claude Code] as the manager and [e.g. Codex] as the worker. Task: add a contact form to my homepage that emails me the message.
Press Enter.
Your agent becomes the planner, reads Zenborg's built-in guide, briefs the manager and worker, and launches the loop. Watch every step live under Loops in Zenborg's top-left menu.
(Already had your AI open before installing Zenborg? Restart that AI window once so it picks up the new tools.)
An agent that grades its own homework misses things.
Example: the February 30th bug
In one of our own test tasks, a calendar module, the worker agent finished building, ran its own self-written tests, saw all green checkmarks, and reported success.
Its logic quietly accepted February 30th as a valid calendar date, alongside two other hidden bugs.
Inside a Zenborg loop, the reviewer caught all three bugs, rejected the draft, and sent exact repair instructions back to the worker. After the repair, the hidden acceptance tests passed.
One example from our own testing, not a guarantee. Results depend on your project, your agents and your tests.
Biggest help: multi-file features and real fixes. Adding a signup flow, wiring up emails, or chasing a bug across five files is where a second pair of eyes catches the most.
Small help: one-line tweaks. Changing a button colour or fixing a typo does not need a three-role team, so there is little to gain there.
Your agent pair matters. Which agents you pick as Manager and Worker changes the result, and a checked loop can take longer on the clock than a solo sprint because it refuses to skip the review.
"I stopped babysitting my agents."
"I built Zenborg because I was spending half my day clicking 'yes' and copy-pasting errors between agent windows.
Once the Manager Loop clicked into place, my agents started checking each other before anything touched my files, and I only reviewed finished diffs.
Now I set up a few projects, let the loops run in their own worktrees, and review the results when I am back."
Let your agents brainstorm with you in one thread.
Stop copying the same prompt between separate agent windows. Bring the agents you already use into one conversation inside Zenborg.
Brainstorm together
Ask for several agents to weigh in on architecture, naming or trade-offs, and read one clean thread.
Direct hand-offs
One agent can @mention another specialist the moment it spots a question outside its lane.
You read one thread. You make the final call.
Need an expert on call? Create one safely in one sentence.
Most people paste random "agent skills" off the internet without checking what is hidden inside. That is how prompt injections and malware sneak into local setups. In Zenborg, you ask your AI for the specialist you want, and it builds a clean, security-checked expert for you:
Use Zenborg. Create a [e.g. SEO copywriter] agent for me: 1. Research online what a great [role] does and which skills matter most. 2. Find the best existing skills for this role and scan every one of them for malware, prompt injections and security holes. 3. Write your own clean version of those skills and install them on the new agent. 4. Give the agent its role: a successful [role] with over 40 years of experience. Write the perfect role description for it.
A few minutes later, your new specialist shows up in every chat and every loop. No config forms. No unvetted third-party files running on your machine.
Start 100% free in public beta. Keep it forever if you love it.
Public Beta ($0 Today)
- ✓Everything unlocked: unlimited local projects, agent loops and custom agents during beta.
- ✓Use your current AI accounts: works with Claude Code, Codex, Hermes, OpenCode and any agent app that supports MCP, under your own accounts and API keys. No markup on AI usage.
- ✓Local-first safety: runs on your own Mac or Windows PC (signed Mac app + official Microsoft Store app). Zenborg never stores or trains on your code.
- ✓Zero trap when beta ends: no credit card required, and nothing is ever billed automatically.
The Optional Founding "Deal Catcher" Licence ($39 Once)
Already know you want Zenborg in your daily toolkit after beta? Lock in the lowest price Zenborg will ever have before our official 1.0 launch moves it to $49:
| What You Get Today | Founding "Deal Catcher" Licence | After Official 1.0 Launch |
|---|---|---|
| One-time price (no sub) | $39 once (or $34 with 3 referrals / $29 with 5 referrals) | $49 once |
| Devices included | 3 personal Macs or Windows PCs | 3 personal devices |
| 1 full year of updates | Clock starts only after beta ends (beta updates free) | Starts at purchase |
| Practical AI Playbook | Included (full copy-paste prompt library) | Included |
| Risk reversal | 30-day unconditional money-back guarantee | 30-day guarantee |
One-time, no subscription: the optional licence is a single payment for the desktop app. AI provider costs are not included and stay between you and your provider.
Want it for $34 or $29? Sign up for the free beta first. Right on the next page, you get your personal invite link: share it with 3 friends to claim $5 off ($34) plus the Bonus Stack, or 5 friends to claim $10 off ($29) plus 3 ready-to-run project templates before you spend a dollar.
Your own AI providers' fees and usage limits still apply. If a worker model hits a limit or fails, you can configure Zenborg to switch to a backup model you choose.
Questions before you start
Do I need to be a programmer?
Which AIs do I need?
Is my code private and safe?
What happens when the beta ends? Will I get charged?
What if the worker writes broken code?
Windows Store user: what if my AI cannot see Zenborg?
How does the $39 Founding "Deal Catcher" Licence work vs. $49, $34, and $29?
Keep your limits. Keep your afternoon. Let the manager handle it.
100% Free Public Beta · No credit card · Mac & Windows · Key in your inbox in under 60 seconds