Your exposure is not the build. It is the Saturday it stops.

By Aghassi Sargsyan · 12 September 2026 · I build this, so read it as interested and check the numbers

What you can put in front of a client

A green tick is not evidence. This is what a run records, and all of it is visible per step rather than summarised at the end:

Source: three runs opened up step by step, including one that failed and one that reported success and was wrong. No account needed, so you can send a client the link before you have built anything.

And Replay re-runs a past execution with its original input against the current definition — so "I fixed it" is something you demonstrate on the exact data that broke, not something you assert.

What it costs, so you can price it

Billed in tokens, not per task, and $29 a month flat for your account rather than per client or per seat. Measured runs:

• A support ticket triaged and answered — $0.0144
• The same job escalating to a human as well — $0.0204
• One that failed on its first tool call — $0.0024
• The dearest agents shipped here — about 11 cents a run

So a client job running 500 times a month is single-digit dollars of usage on top of the $29. Attach your own Anthropic or OpenAI key and the daily ceilings disappear entirely — you pay the provider directly, at cost, with no markup from us. That also means a client's spend can sit on a key you bill through, which is the arrangement most consultants actually want.

The free plan needs no card and covers 100 runs or $0.50 of usage a day, whichever comes first. That is enough to build one real client job and watch it run for a week.

What this does not do for a consultancy

These are the four that would decide it for me if I were you, and none of them is on a roadmap I can promise. They are worth more to you than another feature list.

There is no client separation

Agents and workflows belong to your account. There is no per-client workspace, no client login, no read-only view you can hand over, and no per-client bill. Tags on an agent are the only labelling mechanism there is. If you run five clients, you run five clients' work in one place and keep the map in your head — or you give each client their own account, which is what I would do.

"Which client is quiet?" has no answer

You can see every run and filter by agent. You cannot ask which client's automations have gone silent this week, because there is no client dimension to group by. For a five-client book that is the question you most want answered and it is not answerable today.

Ten tools, not hundreds of connectors

HTTP, webhook, email, Slack, Discord, Google Chat, Telegram, Notion and two Google Calendar tools — plus any REST API you can reach. "Reachable by HTTP" is not the same as a maintained connector with somebody else on the hook when a vendor changes their API. Keep n8n or Make for the wiring.

No SOC 2, no HIPAA

Tenant isolation, JWT auth, rate limiting, secrets encrypted at rest and a GDPR-aware posture — and no certification. Those are audits, not postures, and we do not hold them. If a client's procurement asks, that is a hard stop and no feature comparison changes it.

The failure to design against

This is the one I would want a consultant to know before they put a client's name near an agent. An agent whose lookups fail can still write a confident answer and finish COMPLETED.

Source: a logged run of mine, published in full. Two http_request calls returned HTTP 403 — and the tool result recorded "ok": true, because the request itself completed. The agent read that as data it had. It then told the customer it could see the duplicate charge it had never read. Cost $0.0204, took 28.5 seconds, and nothing about it looked wrong from outside.

So the rule for client work: draft, do not send. Have the agent prepare and a person approve anything that sends, pays or cancels. A wrong draft costs a second of reading. A wrong sent reply costs your client's customer, and then costs you.

Two more worth reading before you rely on a schedule: a scheduled job of mine that logged success hourly for weeks and had never done anything, and 27 branches that were skipped while every run reported COMPLETED. Both are why the skipped-step row in the table above exists.

Trying it on one client job

1. Pick the job that is annoying rather than critical, and that happens weekly.
2. Build one agent that drafts and does not send. Two minutes on the free plan.
3. Run it on twenty real inputs and read the step rows, not the statuses.
4. Show the client one run. That conversation is the actual product.
5. Only then give it a tool that acts, and only one.
6. Put it on a schedule last, and have it report a count of what it processed so a silent week is visible.

Frequently asked questions