Pilata is the control room around it. Plain code claims chats, enforces every guardrail and keeps the journal. Claude does only the judgment — investigating, deciding, writing. Anything it shouldn’t handle goes to a person.
Pilata watches every inbox you connect. Each conversation gets a row in Postgres and a line in the journal. Inboxes you keep for people — WhatsApp, say — stay off-limits in code.
Customers type in bursts. A turn becomes claimable 45 seconds after the last message, so three messages become one job — and one answer.
One row update with SKIP LOCKED. The lease lives only while the turn is worked; a crashed runner simply stops renewing and the turn goes back to the queue.
A fresh context scoped to this one customer. Tools resolve the customer from the conversation — no tool can even name another chat, so nothing leaks between customers.
Before anything leaves, the outbox re-checks: did a person reply since the claim? Quiet hours, reply cap, near-duplicates, card and phone numbers, the inbox’s mode.
In live mode the reply goes out. In approve mode it waits for a tap. If the customer is upset, money is stuck or tools can’t fix it, Claude hands the chat to a manager with findings and the next step.
UPDATE jobs SET state = 'claimed', slot_id = $1 WHERE id = (… FOR UPDATE SKIP LOCKED)
Agents run in an open-source runner on your server, under your own Claude Code login or an API key. The Pilata server does the plumbing and holds the inbox tokens. Neither side ever holds the other’s keys.
Chatwoot inboxes and Telegram team groups. Each source has its own mode — off, shadow, approve or live — enforced by the server.
Queue, dispatcher, timers, outbox guard and an append-only journal on Postgres. It performs every write to a customer — Claude never holds a Chatwoot token.
pilata-runner connect on any box. It dials out over one websocket — jobs come down, tool calls go up. No open ports, works behind NAT.
Pilata replaces one fragile long-running session with a small crew of short, scoped jobs — and a lot of boring, reliable code around them.
Anything with a right answer is a rule you can read. Claude is called when a person would have to think.
Every turn starts clean, scoped to one customer. No compaction drift, no cross-customer leaks — by construction.
Every tool call, draft, send and hand-off — append-only, and live in the UI.
No agent polls. Slots sleep until a turn is ready, and the playbook prefix is prompt-cached.
No mode blocks escalation. The person gets the homework already done.
Возврат 12 400 ₸ по заказу #48190 отклонён банком 2 раза. Карта клиента заблокирована. Дальше: запросить новую карту и повторить возврат.
Every outgoing message — to a customer, to the other side of an order, to a team group — passes the outbox guard. Rules that matter don’t live in a prompt that can be talked around.
Посылка уже в Алматы на сортировке CDEK — обычно это 1–2 дня.
Each source has its own mode, switched in one place. New Telegram groups start in shadow.
Claude works every chat for real, but proposals live only in Pilata’s journal — nothing is written to Chatwoot. Grade them against what your team actually sent.
good for The first week on a new inbox.
“Get the full address”, “write to the seller”, “wait until it’s delivered”. The composer turns a sentence into a task with a goal, required fields, waits and a deadline. Plain code watches the waits; Claude wakes only when one fires.
When knowledge is missing, an agent asks your team’s Telegram group — but first it checks for an open question that already covers it, and joins instead of re-asking. One answer resumes every waiting chat, with per-chat exceptions in the same line.
Можно вернуть бонусы за отменённый заказ, если он был оплачен бонусами?
Feedback on a chat, a note in the Teach box or /teach in Telegram becomes one short rule. Claude drafts it, merges it with lessons on the same situation and flags contradictions; a person approves it. From then on it is in every prompt — in full.
Our first customer, a second-hand marketplace, ran Pilata in shadow on its main inbox: Claude worked every chat, nothing was sent, and every proposal was graded against what the human team actually replied.
The runner is a small open-source CLI. Log in to Claude Code on your server — or set an API key — and connect. Add slots for more parallel chats, add runners for more machines. Each slot runs one short job at a time on the Claude Agent SDK.
Start in shadow on one inbox. Go live when the grades say so.