Telegram · Slack · Discord · Web · Bring your own model
Bots that answer your customers from your own documents — and a developer bot that can read logs, query databases and run deploys, where nothing destructive happens until a named person approves the exact command.
why is checkout failing for EU customers this morning?
The EU gateway rejects payments when the billing country and the card BIN disagree. Your runbook says to re-run the reconciliation job before retrying.
ok restart the payments worker on prod
Only Marco can approve. Expires in 5 minutes — nothing runs.
Upload PDFs, Word files, Markdown, or point it at a page. Every bot gets its own isolated RAG knowledge base — retrieval-augmented generation, so answers are grounded in your documents and cited rather than guessed. One bot can never retrieve another’s passages. When nothing relevant is found it says so instead of inventing an answer, and you can read and delete the individual chunks it drew on.
Read logs over SSH, query a database, check a repository, call a REST API you registered, run a deploy script you pre-approved. Tools are attached per bot in the execution layer, so a bot never granted one cannot use it however the conversation goes.
A customer can ask for a person, or the bot escalates itself when its confidence drops. The conversation appears in a live inbox, the bot pauses mid-thread, and your agent replies through the same channel the customer is already in. Close it and the bot picks up where it left off.
Tools are what a bot may touch. Skills are named workflows you write once and invoke with a slash. Both are attached per bot, so two bots in the same workspace can have completely different reach.
Connecting a new app is configuration, not a release. You register the address, the login and which actions count as reading — the bot never sees a password or a key.
⚠ A skill is text and grants nothing. It cannot add a tool, widen a scope, reach a system the bot was not given, or skip an approval.
The open-source way to get here is a weekend: a machine, a database, a vector extension, a queue, a reverse proxy, certificates, and a model runtime you keep feeding. Then you own all of it forever.
Subscribe, choose the address your team signs in at, paste one bot token. Everything above is already running, already backed up, already patched — and it is our pager, not yours.
The hard part of this product was never the bots. It is giving one write access to real systems without that being a bad idea — so the safety stack is a subsystem, not a paragraph in a system prompt.
Deterministic, and not an LLM. It force-flags any non-SELECT statement, any write, delete or deploy pattern, and anything aimed at production. The model may add a flag. It can never remove one.
The allowlist lives in the execution layer, outside the model, so an out-of-scope request has no tool to call even if the prompt is subverted. A workflow you write is text: it cannot add a tool, widen a scope, or skip a signature.
The approval shows the literal command, never a paraphrase — approving a summary is approving something other than what runs. Only the person who asked can sign it, it expires in minutes, and silence is a refusal rather than a default.
Every tool run — allowed, blocked, or never answered — lands in an append-only log with the exact command, the environment, and both verdicts. The table rejects UPDATE and DELETE at the database, so it is append-only in fact rather than by convention.
Payment happens on Stripe’s own page. Nothing card-shaped reaches our servers.
Row-level security and a database role per customer — not a WHERE clause we remembered.
Every tool run written down. The table refuses UPDATE and DELETE.
Same persona, same documents, same spending caps, same handover inbox. Connecting a second channel changes nothing about the first.
Direct messages and groups. Answers a mention or a reply.
Runs toolsReplies in a thread under whatever it is answering, inside your own workspace.
Runs toolsAnswers /ask in your server, under your own bot name and avatar.
Runs toolsA chat widget you add with one script tag. Origins are allow-listed per site.
Answers onlyEvery plan carries a ceiling the platform enforces itself. Reach it and the bots pause. There is no route to a surprise invoice, because there is no overage you did not deliberately switch on.
One bot, for a small support inbox.
More bots and more room, billed if you go over.
Unlimited usage for a busy team.
Nothing to sign up for elsewhere. Your plan’s monthly tokens cover every bot, every channel and every tool call, and the platform pauses at the ceiling rather than billing past it.
Best if you want one invoice and no accounts to manage.
Connect your own provider key and those tokens are billed by them, directly, at their rates — so nothing you spend counts against a plan quota and nothing pauses at a ceiling. You still get the bots, the channels, the tools, the inbox and the audit log.
Set per bot, so a customer-facing bot can run on ours while the developer bot runs on yours.
Prices excluding tax. Your key is encrypted under your workspace’s own key and is never sent anywhere but the provider you registered it for.
Choose a plan and the address your team signs in at. The first bot can be answering before the end of the meeting you are supposed to be in.