Half-built and forgotten
A dozen flows built over a weekend. A few months later, most are broken.
HomlaAgents
Helpers that do the boring jobs, and ask first.
Start from a template or describe the job in plain words. Pick its tools. Choose what needs your yes. It runs on Claude or Codex.
AgentsNew agent
CancelCreate agent
Start from
A/B test watcherInbox triageWeekly client reportBlank
The job
Each morning, sort the support inbox. Draft replies for order questions using the Refund reply skill. Never promise a refund over $200.
Tools and approvals Each write asks you first unless set to Auto.
outlook_mail_searchReadAlways allowed- outlook_draft_createWriteAutoAsk first
- outlook_mail_sendWriteAutoAsk first
shopify_orders_listReadAlways allowed
Who can talk to it
Everyone in HomlaSpecific peopleAnyone
Support (role)[email protected]
When it runs
- On a scheduleWeekdays, 08:00
- When an email arrives
- Only by hand, with Run now
Limits and engine
- Engine
- Claude · your plan or your Anthropic key
- A run
- 10 minutes at most
- A day
- 5 automatic runs at most
The problem
Agents get built, then left in draft.
Nobody owns them. They don't fit the real work. Nobody trusts them alone. So they sit half-built, or someone checks every output by hand.
Nobody owns it
No owner, no clear job. The results land somewhere nobody opens.
Too risky to let it act
So you keep it in draft and read every message before it goes out.
What you control
You set the job, the tools and the people.
Its tools, and what needs a yes
Pick the tools it may use. Looking things up is always allowed. Each change asks you first, unless you set that tool to Auto.
- Approve, edit then approve, or reject
- Requests expire if nobody answers
Who can talk to it
Your whole team, chosen people, or anyone. Block anyone you want. Requests from outside your team always wait for approval.
- Roles decide who can use or manage agents
Runs when you want
On a schedule, when an email arrives, or by hand. Every run is kept step by step.
- Its own email address
- Runs on a teammate's computer or a cloud box
- Limits per run and per day
Also in Homla
The rest of Homla.
Can an agent send something without asking?
Not unless you allow it. Each change waits for a yes. Set a tool to Auto once you trust it.
Where does it run?
On a teammate's computer while it is on. Or on an always-on cloud box from $25 a month.
Does an agent take a seat?
No. Agents are not seats. You only pay for a cloud box if you want one.
What if a stranger emails it?
What comes in is read as information, never as orders. Any change a stranger asks for waits for your yes.
Hand off the boring jobs. Keep the final say.
Start free. Pick a template. Approve its first change yourself.
- Playbooks the whole team can use
- One memory every AI reads
- Each person gets only their access
- Works in Claude, ChatGPT and Cursor