Emissario
PT Ask the AI
AI assistant

An assistant that answers your visitors at any hour.

It talks with the people who visit your website, answers from the information you gave it and shows your team what people want. You decide what it knows, how much it may spend and what it may look up.

Talk to this site's assistant Get it set up for you
01 · What it does

A front desk that never takes a day off, and follows your rules.

Answers at any hour

Questions about your services, prices, opening hours or your work, answered right away, at night and on weekends too.

Uses your information

It follows the texts you wrote and looks up your other documents when a question calls for it.

Stays within your budget

Daily usage limits you set: AI spending never goes past the cap.

Tells you what people want

Anonymous summaries of conversations, when the visitor chooses to send one to your team.

Is honest with visitors

It says it is an AI, does not pretend to be a person and does not reveal its internal instructions.

Looks up current data

If you allow it, it checks a price list or another page of yours to answer with today's information.

02 · What it knows

You teach the assistant with plain texts.

Everything it knows comes from documents you write, as if you were briefing a new receptionist: who you are, what you offer, prices, opening hours, common questions and the tone to use.

  • Write them yourself, or explain it all in a conversation and, with the set-up service, the texts are written for you.
  • Changed a price? Update the text, and the assistant answers with the new information.
  • Long documents, such as a price table or a returns policy, are only read when a question calls for them.
  • The texts are kept on GitHub, with the history of every change.
the assistant's instructions (example)

Example Consulting's assistant

You answer the visitors of Example Consulting's website, in London.

  • Answer in English, in a friendly and concise way.
  • We are open Monday to Friday, 9 am to 6 pm.
  • To talk about prices, read the prices document.
  • To book a meeting, point to the contact form.
  • Do not promise deadlines: say the team confirms by email.
03 · Spending under control

You decide how much it may spend.

AI providers bill by the amount of text the AI reads and writes, counted in “tokens” (pieces of words). Emissario checks the limits before each answer, so a busy day never goes past the cap you set.

Length of each answerA maximum per answer, to avoid long and costly replies.
Daily cap for the siteWhat all visitors together may use in a day.
Daily cap per visitorNo single visitor can use up the day's budget.
Messages per minuteSlows down anyone who sends too many messages.
  • When a limit is reached, the visitor is asked to come back later, and the chat opens again at midnight.
  • The dashboard shows each day's usage next to the limit, and who reached their own limit.
Dashboard · AI usage AI usage on the dashboard: the spend of each day and of each visitor, and who reached the daily limit
AI usageToday's spend, each day's and each visitor's, and who reached the limit. Demo data.
04 · Conversation summaries

Learn what visitors ask, without learning who they are.

At the end of a conversation, the visitor can send a summary to your team. That is how you find out the most common questions, the requests for quotes and what your website is missing.

  1. 01

    The visitor chooses to send

    A button in the chat, after an answer. Nothing is sent unless they ask.

  2. 02

    The AI writes a summary

    Without names, dates, places, documents, phone numbers, emails or addresses.

  3. 03

    An automatic check cleans up the rest

    Whatever still looks like an email, a phone or a card number is replaced by a label.

  4. 04

    Your team reads it on the dashboard

    Stored encrypted, without the visitor's internet address, for 90 days by default.

the automatic check, illustrated

Intent: pricing. A visitor asked for a quote and to be contacted at ana@example.com [email] or +1 202 555 0100 [phone].

  • Search the summaries by date or by word.
  • Summaries stay off until you turn them on, after a notice on the personal data they may hold.
  • An AI can make mistakes, so a summary is protected as personal data and can be erased on request.
  • This site's chat has the Send a summary to the team button: try it.
Dashboard · summaries A list of anonymous conversation summaries
SummariesSearched by date and by word, with no visitor's address.
Dashboard · a summary An anonymous conversation summary, with its date and the visitor's intent
A summaryWhat the visitor wanted and whether they got an answer.
05 · Transparency

Honest with visitors, by default.

It says it is an AI

In its first answer and whenever asked, and it never poses as a person. The chat also says so before the first message, as the EU AI Act asks.

The conversation stays with the visitor

It is kept in the visitor's browser, not on the server, with a digital signature: nobody can forge an answer from the assistant.

It keeps your secrets

It never reveals its internal instructions, passwords or system addresses, and does not obey orders hidden in the documents it reads.

What the AI provider receives

The AI provider is another company, which receives the conversation to write each answer. Sign a data processing agreement with it, or use a model on your own server, with Ollama, so that conversations never leave your infrastructure.

SentWith each message
The conversationEvery message of the visitor and of the assistant, as typed.
The instructionsEmissario's rules and your instructions.
What it looked upThe documents it read and what the pages it checked answered.
A summary requestThe whole conversation again, only when the visitor asks.
06 · Which AI to use

Pick the AI you prefer.

The assistant works with several AI providers. You use your own account and pay them directly, or you run a model on your own server.

  • OpenAI
  • OpenRouter
  • Groq
  • Ollama (on your server)
  • The key to your provider account stays hidden: the dashboard shows at most its last four characters.
  • With Ollama, conversations never leave your infrastructure and there is no per-message cost.
Dashboard · assistant settings Assistant settings: the AI provider, the hidden key, the model and the spending limits
Assistant settingsThe provider, the hidden key, the model and the limits. Here it answers through a local test provider.
07 · Look-ups

Let it look things up, only where you allow.

You can allow the assistant to check specific addresses, such as a price list or a calendar of open slots, to answer with up-to-date data.

  • It only checks what you registered, by name. Never an address it made up.
  • The passwords to those addresses stay hidden, even from the AI model itself.

Anything the assistant can look up, any visitor can ask it about. Only allow information that may be public.

For the technical team

Technical details.

How the assistant works inside, for whoever evaluates or installs it. The agent documentation has everything.

The path of a message
  1. The browser sends the whole conversation and the signature of the last answer to POST /api/agent/messages. The history stays in the visitor's browser.
  2. Emissario checks the origin, the IP blocks, the signature, the requests per minute and the tokens left today, before calling any model.
  3. The rules come in: Emissario's own first, then your instructions/README.md. No message can change them.
  4. Tools: the model can list and read the files of instructions/, or call the declared URLs, in up to 5 rounds per message.
  5. The answer comes back signed with HMAC-SHA256. Emissario records the tokens used, never the conversation.
The instructions repository
  • The agent repository is synced like the website: on a schedule, or with the Sync button. The files are read on every message.
  • instructions/README.md holds the system instructions. Without it, the chat answers 503 no-instructions.
  • instructions/SUMMARY.md, optional, adds rules for summaries, such as classifying the intent.
  • The model lists and reads the other files as it needs them, up to 64 KiB each. Hidden files are left out.
  • emissario drop-agents-skills writes the skill Claude Code, Codex and Antigravity load to write these files.
my-agent/
└── instructions/
    ├── README.md      # who the agent is, what it answers
    ├── SUMMARY.md     # optional: rules for summaries
    ├── pricing.md     # read when a visitor asks
    └── policies/
        └── refunds.md
Limits and defaults
SettingWhat it limits
max-tokens (default 1024)Output tokens of each call to the provider.
daily-limitInput and output tokens of the day, all visitors together.
ip-daily-limitThe tokens of one IP in a day; an IPv6 address counts by its /64.
rate-per-minute (default 10)Requests per minute from one IP.
  • A call in flight reserves max-tokens until it answers, so concurrent requests never pass a limit together.
  • Every tool round and every summary counts, input and output.
  • Over a limit: 429 daily-limit or ip-daily-limit, until midnight.
  • A conversation: up to 50 messages, 4,000 characters per message, 64 KiB per request, 60 s per answer with tools.
  • Usage is kept for 90 days, with the IP hashed by default; a viewer sees IPs masked.
Providers and the API key

Any model compatible with the OpenAI API that supports tool calling. Emissario calls <base-url>/chat/completions.

ProviderBase URL
OpenAIhttps://api.openai.com/v1
OpenRouterhttps://openrouter.ai/api/v1
Groqhttps://api.groq.com/openai/v1
Ollama (local)http://localhost:11434/v1
  • The API key is write-only: a page shows at most its last four characters, never to a viewer, never in a log.
  • EMISSARIO_AGENT_API_KEY takes precedence over the key saved in the backoffice.
Summaries
  • The chat sends the conversation and the signature to POST /api/agent/summaries; a conversation the agent never had is refused.
  • The second pass replaces a CPF, a CNPJ, an email, a phone and a card number with [cpf], [cnpj], [email], [phone] and [card].
  • The visitor gets a 16-character reference. A root deletes one summary, or all of them before a date, and can put one on legal hold.
URLs the agent may call
  • Each URL has a name, a description, GET or POST, and headers the server itself adds.
  • The model never writes a URL: it calls an endpoint by its name.
  • Headers are write-only: never shown to the model, to a viewer, or again to a root.
  • No redirect is followed; only text or JSON is read, cut at 32 KiB, within 10 seconds.
name
implementation-prices
method
GET
url
https://example.com/prices.json
description
The prices of the plans. Call it before talking about prices.
headers
Authorization: ••••••••
Try it

Talk to it. Then get your own.

This site's chat runs on Emissario. Ask it anything, then see how to have an assistant like it on your website.

Ask the AI Get it set up for you