Carpe Lab

Grow your business without growing your headcount.

AI Employees work inside your business. You message them like a colleague, they take work off your plate, and they get better every week. Like giving everyone on your team three assistants.

See how we work

Works in Slack Teams WhatsApp Email

Hermes agents, hosted properly.

Run as many AI employees as you need. Each one isolated, every tool call permissioned and logged, and a Console that shows every change they make.

Take the stack apart

Connects to Gmail Calendar Xero HubSpot Slack+1,367 out of the box · custom APIs · managed browser

Runtime · HermesIsolation · Per agentCredentials · Never in the agentAudit · Every call

Max · AI Employee online in #sales

Chase the quiet proposals

Four proposals with no reply in over a week. Three are under £5k, one is Northwind at £18k. Shall I nudge all four, or leave Northwind for you to call?

Nudge the three, I'll call Northwind

Sent, in your usual tone. Northwind is in your calendar for 2pm with the proposal attached, and I'll tell you who replies.

Write up Acme's weekly report

Spend was £4,200 of £4,500, 61 leads, cost per lead down 12% on last week. The retargeting ad is starting to fatigue. Client version or the internal one?

Client version, send it to Sam

Sent to Sam with the three charts, and a copy is in #acme. I suggested fresh creative for retargeting next week.

What's on today?

Three calls. A demo with Brightline at 10, Acme's monthly review at 1, and Halden's renewal at 4. Halden opened the pricing page twice this week. Want a one-pager before the call?

Yes please

Done, it's attached to the invite. Brightline's deck is in the thread, and Acme's review notes are ready to go.

Max is ready when you are.

Give Max a job:

That was one message. Now picture your whole team.

Meet your AI Employee

Message it like you'd message a colleague.

Your AI Employee has its own name and works where your team already talks. No new app to learn. No training course.

Like a good new starter, it checks with you when it's not sure. And when you correct it, it remembers. You teach it in plain English. No tech skills needed.

Press the button. Everyone gets three.

Slack Teams WhatsApp Email

Jess: Chase the quiet proposals

JessSalesLeadsFollow-upsDiscounts
PriyaMarketingCampaignsReportsBudgets
TomAccountsUpdatesRenewalsInvoices
DanFounderNumbersApprovalsBudgets

Asking is the easy part. The work lives in your tools.

It works with what you already use

If your team can use it, your AI Employee can too.Connect to anything. Three ways in.

1,367 tools work out of the box. Your own systems, we connect. And anything with a login page, it can use through a browser, the same way a person would.

1
1,367 tools, ready to go1,367 connectors via OpenConnect
Email, calendars, CRMs, ad platforms, analytics, messaging. Connected in minutes.
OAuth handled for you. Tokens live in the gateway's vault, never in the agent.
and 1,363 more
2
Your own systemsCustom connectors, written for your API
Your CRM's custom fields, the proposal tool, the spreadsheet the pipeline really runs on. We write the connection.
Typed MCP tools over your API, with writes allowlisted one by one. The agent sees the tools, never the keys.
3
Anything with a login pageA managed browser, for everything else
Partner portals, ad platforms with no API, the client's old dashboard. Your employee signs in and uses it like a person. This is the big one.
Headless browser sessions with sign-in handled by the platform. Every action is a tool call: permissioned, logged, and visible in the Console.

adportal.example/loginmanaged browser

No API? No problem. It signs in and uses the screen, like a person would.

Once it's connected, it just gets on with the day.

A day with your AI Employee

Work gets done while your team gets on with theirs.Every step, timed and logged.

Scroll through one ordinary day. Nothing here is a demo flow. These are the jobs an AI Employee picks up around a real sales and marketing team.

07:00keep scrolling to move through the day

07:00
Pipeline report postedcron.weekly → crm.pipeline → slack.post
Every Monday, from the CRM, in #sales before anyone is in.
2.4s
08:12
New lead landedform.inbound → crm.create → email.send → slack.post
A form fill became a lead in the CRM, got a reply, and a card in Slack, in 40 seconds.
600ms
09:40
Demo booked from a replyemail.inbound → calendar.book → email.send
The lead replied “Thursday works”. Calendar updated, invite and reminder sent.
450ms
11:05
Asked before guessingdeals.discount → denied → ask(jess@) · 31s wait
A lead asked for 15% off. It checked with Jess instead of making a call on price.
31.0s
14:02
Proposal sentemail.send → allowed · 1,204 tok
The agreed price, with the case study closest to their industry attached.
300ms
16:15
Client report sentads.report → docs.render → email.send
Spend, leads and cost per lead for the week, with charts, in the client's format.
1.8s
17:30
Chased the quiet onescrm.deals stale=7d → email.send ×3
Three proposals with no reply in a week got a friendly nudge.
900ms

A few of the jobs it can pick up:

Answer every new leadChase every proposalKeep your CRM up to dateBook demos from repliesSend client reports on timePrep you for every callRe-engage cold leadsTurn a call into a proposal

That was one turn. Here's what ran it.

Runtime · Isolation

Every agent in its own box.

Each AI employee is a Hermes agent (open source, from Nous Research) running in its own isolated container, with its own filesystem, memory, scheduled jobs and skills.

One agent can't see into another, and clients are kept apart. Prefer Claude Code, OpenClaw or another agent? They run on the same platform, behind the same gateway.

Fast deploys

New agents and skill changes go live in minutes.

Kept up to date

We track upstream Hermes releases so you don't.

Backups

Memory, skills and state backed up on a schedule.

Rollback

Every release is a pinned image. Roll back with one command.

Now take the whole thing apart.

Architecture · scroll down through it

One stack. Take it apart.

Five layers between a message in Slack and a change in your CRM. Every hop is authenticated, checked and logged.

SSO and their own permissions. The agent can only do what the person asking could do.

Every call in that stack goes through one place.

Gateway · Permissions

One gateway for every tool.

Every tool call from every agent goes through one MCP gateway built for your business. It is the only thing that can reach your systems.

Limits live in code, not in prompts, so an agent can't talk its way past them. Every allow and every deny is logged.

gatewayper person, per tool

crm.write
crm.read
email.send
deals.discount
ads.read
analytics.read
ads.budget
finance.read
finance.*

green = allowed · orange = denied · marketing can't move ad budgets, sales can't give discounts

No credentials in the agent

API keys and OAuth tokens live behind the gateway, one connector per system.

Checks the person, not the bot

Every call is authorised with OpenFGA against the human who asked, per person and per tool.

Reads broad, writes explicit

Agents can look things up widely. Anything that changes data has to be allowed.

Guard rails in code

Email with no delete. Slack in allowed channels only. Discounts and ad budgets need a human.

It works with what you already use

If your team can use it, your AI Employee can too.Connect to anything. Three ways in.

1,367 tools work out of the box. Your own systems, we connect. And anything with a login page, it can use through a browser, the same way a person would.

1
1,367 tools, ready to go1,367 connectors via OpenConnect
Email, calendars, CRMs, ad platforms, analytics, messaging. Connected in minutes.
OAuth handled for you. Tokens live in the gateway's vault, never in the agent.
and 1,363 more
2
Your own systemsCustom connectors, written for your API
Your CRM's custom fields, the proposal tool, the spreadsheet the pipeline really runs on. We write the connection.
Typed MCP tools over your API, with writes allowlisted one by one. The agent sees the tools, never the keys.
3
Anything with a login pageA managed browser, for everything else
Partner portals, ad platforms with no API, the client's old dashboard. Your employee signs in and uses it like a person. This is the big one.
Headless browser sessions with sign-in handled by the platform. Every action is a tool call: permissioned, logged, and visible in the Console.

adportal.example/loginmanaged browser

No API? No problem. It signs in and uses the screen, like a person would.

Everything it does leaves a trace.

A day with your AI Employee

Work gets done while your team gets on with theirs.Every step, timed and logged.

Scroll through one ordinary day. Nothing here is a demo flow. These are the jobs an AI Employee picks up around a real sales and marketing team.

07:00keep scrolling to move through the day

07:00
Pipeline report postedcron.weekly → crm.pipeline → slack.post
Every Monday, from the CRM, in #sales before anyone is in.
2.4s
08:12
New lead landedform.inbound → crm.create → email.send → slack.post
A form fill became a lead in the CRM, got a reply, and a card in Slack, in 40 seconds.
600ms
09:40
Demo booked from a replyemail.inbound → calendar.book → email.send
The lead replied “Thursday works”. Calendar updated, invite and reminder sent.
450ms
11:05
Asked before guessingdeals.discount → denied → ask(jess@) · 31s wait
A lead asked for 15% off. It checked with Jess instead of making a call on price.
31.0s
14:02
Proposal sentemail.send → allowed · 1,204 tok
The agreed price, with the case study closest to their industry attached.
300ms
16:15
Client report sentads.report → docs.render → email.send
Spend, leads and cost per lead for the week, with charts, in the client's format.
1.8s
17:30
Chased the quiet onescrm.deals stale=7d → email.send ×3
Three proposals with no reply in a week got a friendly nudge.
900ms

Observability

See everything they do.

One OpenTelemetry trace per agent turn, plus logs and alerting. Traces carry metadata only, never message text, tool arguments or results.

The Console shows every conversation down to each tool call, skills versioned with diffs and who changed what, the rules your staff have taught it, its memory and scheduled jobs, and token spend per agent and per person.

Observability runs on a self-hosted stack per client, not a third-party service, because it's your data.

console / sales / conversationstoday

crm.search_contactsjess@allowed
crm.list_dealsjess@allowed
email.sendjess@allowed
deals.apply_discountjess@denied

skills / send-proposal.md · edited by jess@ from a chat

- Send the proposal to the lead.

+ Attach the case study closest to their industry. Never offer a discount.

So how do you get one?

Who decides what it can touch? You do.

How we work with you

We sit down with your team and train your new employee.

You know your business. We know what AI Employees can do, and how to train them so they actually work.

Put the two together and everyone on your team gets superpowers. Assistants that understand their part of the business, and how all the messy bits connect.

1
We listen
We sit down with you and your team and find the work everyone would love to get off their plate.
2
We connect it
We hook your new employee up to the tools you already use. Your inbox, calendar, CRM and ad accounts.
3
We train it
We write its playbook from your know-how, then test it on real work with your team until it gets it right.
4
It keeps getting better
Your team teaches it as they go, and we stay on hand to add new jobs as you think of them.

Control

You stay in control.

Run as many agents as you need, across teams, without losing track of what each one can touch.

Single sign-on

Your team signs in with the identity provider you already use.

Per-person access

Decide who can ask which agent to do what.

Staging agent

Test skill and config changes before they reach your team.

Self-hosted

Run the same stack on your own infrastructure, with full control. Coming soon

Who you'll work with

Built by builders.

We're two software engineers with nearly a decade each, from startups to huge enterprises, including Meta and Vercel. We're good at taking messy business problems and solving them.

And we'll tell you plainly what's useful, and what to ignore. You'll talk to us, not a sales team.

We run this platform ourselves, every day, for real businesses. You'll talk to the people who built it.

Getting started

Two ways to work with us.

Pick how much we run for you.

Setup

We do the groundwork first, then spend a day with you getting your AI Employee set up and connected to your tools. After that we host it and keep it running for a monthly fee.

Full engagement

We map out how your business runs with you and your team, connect your employee to all of your systems, and keep training it with you month after month.

Managed setup

One day with your team. Agents deployed, connected to your tools through the gateway, then hosted, monitored and kept up to date by us.

Full engagement

Process mapping across the business, custom connectors for your systems, and ongoing skill development with your team.

Self-hosted Coming soon

The same stack on your own infrastructure. You run it, with full control.

Not sure which fits? Book a call and we'll tell you straight.

Grow your business without growing your headcount.

Hermes agents, hosted properly.

Message it like you'd message a colleague.

If your team can use it, your AI Employee can too.Connect to anything. Three ways in.

Work gets done while your team gets on with theirs.Every step, timed and logged.

Pipeline report postedcron.weekly → crm.pipeline → slack.post

New lead landedform.inbound → crm.create → email.send → slack.post

Demo booked from a replyemail.inbound → calendar.book → email.send

Asked before guessingdeals.discount → denied → ask(jess@) · 31s wait

Proposal sentemail.send → allowed · 1,204 tok

Client report sentads.report → docs.render → email.send

Chased the quiet onescrm.deals stale=7d → email.send ×3

Every agent in its own box.

Fast deploys

Kept up to date

Backups

Rollback

One stack. Take it apart.

One gateway for every tool.

No credentials in the agent

Checks the person, not the bot

Reads broad, writes explicit

Guard rails in code

If your team can use it, your AI Employee can too.Connect to anything. Three ways in.

Work gets done while your team gets on with theirs.Every step, timed and logged.

Pipeline report postedcron.weekly → crm.pipeline → slack.post

New lead landedform.inbound → crm.create → email.send → slack.post

Demo booked from a replyemail.inbound → calendar.book → email.send

Asked before guessingdeals.discount → denied → ask(jess@) · 31s wait

Proposal sentemail.send → allowed · 1,204 tok

Client report sentads.report → docs.render → email.send

Chased the quiet onescrm.deals stale=7d → email.send ×3

See everything they do.

We sit down with your team and train your new employee.

We listen

We connect it

We train it

It keeps getting better

You stay in control.

Single sign-on

Per-person access

Staging agent

Self-hosted

Built by builders.

Two ways to work with us.

Pick how much we run for you.

Setup

Full engagement

Managed setup

Full engagement

Self-hosted Coming soon