Agentyard — by hotmun
A team that answers customers on the channels they already use, works to a schedule, and keeps your site live — on isolated infrastructure you can host yourself.
Every business gets its own pod and its own database. Nine messaging channels, 27 model providers, and spend ceilings that hold. No free tier — a trial, then a plan.
New message while the van was on a job — boiler leaking under the sink, can anyone come tonight?
Replied in 11 seconds, asked for a photo and a postcode, logged the lead.
Priced against the rate card, sent a signed quote link. No customer account needed.
Quote accepted. Booked 07:30 Wednesday, owner notified, calendar updated.
Scheduled turn skipped — no social account connected yet. Nothing spent.
You describe the work. Each agent gets a role, the tools for it, and a budget. Below each promise is the mechanism that makes it hold.
01
Customers write where they already are — WhatsApp, Telegram, Slack, Discord, email, iMessage, Matrix, Lark, IRC. An agent picks it up, asks the right follow-up, and writes the lead down instead of letting it sit in an unread inbox.
Every channel is declared once in a single manifest, so adding credentials hot-reloads the connection — no restart, no redeploy.
02
Chase unpaid invoices on Mondays, post on Thursdays, follow up two days after every quote. Cadences run whether or not anyone opened the app.
A scheduled turn is gated before it costs anything: if a prerequisite is missing — no connected account, no key, a provider parked — the turn is skipped at zero model spend and re-checked on the next tick, instead of burning a call to rediscover the same wall.
03
Ask for a site and you get one — pages, photos, and a form that actually reaches you. Enquiries land in your business's own database and you get told about them.
Static pages are served from a CDN, so visitors never wait on the application, and a restart is not an outage. Form posts go to the database inside your pod — business data never moves to a shared one.
04
Agents keep their own memory and search it before answering, so a returning customer isn't asked the same question twice. Work moves between agents like it would between colleagues.
Per-agent storage with full-text and vector search, recalled into context on each turn. A task that goes two turns without progress escalates — assignee, then team lead, then the executive, then it waits for you rather than looping quietly.
05
The failure mode of AI agents is a quiet runaway bill. This one has a ceiling per hour and per day, and it is enforced before a call is made, not discovered on an invoice.
You're warned at 80% of a ceiling; at the ceiling, turns defer
instead of retrying. If every model provider is failing, work is
held with a reason attached and released the moment the wall
clears — waiting_on, not a blind retry timer.
What you do, who writes to you, what a job costs. You get a team with roles that fit — reception, estimating, scheduling, marketing — not a blank prompt box.
Paste one credential and the connection comes up live. Start with the inbox that actually rings; add the rest whenever you like.
Pick what the team may spend a day. Every reply, scheduled turn and pound spent is on one screen — and you can stop the whole team from it.
Multi-tenant usually means your records sit in a table next to someone
else's, separated by a WHERE clause. Here each business
runs as its own workload with its own database and its own storage
scope, because the boundary is the product.
The same binary you'd run yourself is the one we host. Bring your own model keys and the spend is yours; use ours and the ceiling is still yours. Nothing here needs a hyperscaler — it runs on our own cluster, and it will run on yours.
One credential brings a channel up live — no restart. The same agent answers across all of them, with the conversation kept in one place.
Connect one channel, give the team one job, and see what a week looks like when nothing goes unanswered.
Trial included, then a plan — there is no free tier, because a team that runs all day costs real money to run.