An AI operator answers your customers. You decide how far it goes.

It answers from your company's documents, stops to ask you before doing what you told it not to, and turns off with one button.

Join the waitlistAsk us to build it
shipping-assistantv1On shiftRecorded run

What your customer sees

  1. Customer

    If I order today, how long does delivery to Córdoba take?

  2. Agent

    Delivery to Córdoba typically takes 3 to 5 business days since it is outside the warehouse's city [1]. We dispatch orders Monday to Friday [1].

    SourceShipping and delivery times doc-shipping v3

  3. Customer

    I ordered yesterday, order 48213. When will it reach Córdoba?

  4. Agent

    Regarding your question about delivery to Córdoba, we dispatch from Monday to Friday [1]. Orders confirmed before 1 pm leave the same day [1]. Since Córdoba is outside the warehouse's city, your parcel should arrive in 3 to 5 business days [1]. Please note that specific confirmation for a numbered order requires checking with a person on the team [2].

    SourceShipping and delivery times doc-shipping v3Changing a delivery address doc-address v2

Pick what the customer writes

In this recorded run the customer writes one of these two messages.

What you see

Where it is

  1. Reads the message
  2. Searches your documents
  3. Writes a reply2 times
  4. Checks it against your rules2 times
  5. Asks you
  6. Replies
  7. Finishes

The shift owner approved. The draft went out.

Recorded on October 1, 2026 on Line's real harness with a free local model; the documents live in memory. Played back faster than it happened.

An agent is what you write about it.

The one above is this and nothing more. Someone in operations builds it in the browser, without code, in any order they like.

Spec
Its job, what it reads and what it never does
The agent above
goal
Answer customers' questions about shipping and delivery in English, only from the team's documents.
reads
the order number, the delivery city
never
  • give a delivery date for a numbered order: a person on the team confirms it
  • ask for passwords or personal details that are not needed
handoff
A person on the team is checking your order and will write to you today before 6 pm.
Graph
The steps it takes with each message
The agent above

Seven steps from message to reply, with the check before anything is sent.

Reads the message, Searches your documents, Writes a reply, Checks it against your rules, Asks you, Replies, Finishes

Tools
What it may touch
The agent above

Search and read in the Corpus, reply with a source, ask you.

corpus.search corpus.get_document reply escalate

Evals
Your test questions. Each run shows how many cases each scorer met.
The agent above

None: it is a sample made for this page and never went to production.

Launch
Which version is on and who answers for it
The agent above

Version 1.

Five parts. One launch.

Each part of an agent is its own layer in Line. Watch them fill in, one feeding the next, until the first run starts.

Illustration with a sample agent. Not a recording.
  1. Write the SpecIts goal, what it reads, what it never does.
  2. The Graph fills inThe steps every message goes through.
  3. Connect the toolsOnly what the harness declares, and a Corpus key.
  4. Run the evalsYour test questions run as real conversations. Each scorer reports how many cases it met.
  5. LaunchYou read the latest run and put a version on shift.
  6. The first runA customer writes. The route lights up, step by step.

It may read and reply. For anything else, it waits for you.

What it can do lives in its tools, not in an instruction it could ignore.

It may

  • Search the documents you gave itcorpus.search
  • Read one of those documents in fullcorpus.get_document
  • Reply, always citing a documentreply
  • Stop and hand you the conversationescalate

It waits for you when

  • Its reply breaks a rule or lacks support, twice in a rowcheck_failed
  • Your documents say nothing about what it was askednothing_retrieved
  • It reaches the cap on steps or text for its turnlimit_reached
  • You took away its access to your documentscorpus_refused

A reply without a source does not go out: the tool refuses it and the refusal is recorded.

The off switch is yours.

Stop cuts all its keys at once. No new conversation gets in until you launch it again.

Stop
Its keys stop working on the spot: anyone who writes to it is told the agent is stopped.
Shift owner
One person per agent on shift. When the agent stops to ask, they get an email.
Versions
Every change is a new version, tested before it goes out. To go back, you launch an earlier one.

If you write code, it's an API.

Your system opens a conversation with the agent's key and reads each turn as events, step by step. The widget for your site comes later.

The API is built; it opens when Line opens.

Open a conversation
curl -X POST https://api.line.trysenda.xyz/v1/conversations \
  -H "Authorization: Bearer $LINE_AGENT_KEY"
Send a message and read the turn
curl -N -X POST https://api.line.trysenda.xyz/v1/conversations/$CONVERSATION_ID/messages \
  -H "Authorization: Bearer $LINE_AGENT_KEY" \
  -H "Content-Type: application/json" \
  -d '{"text":"If I order today, how long does delivery to Córdoba take?"}'
What the run above returned
id: 1
event: step
data: {"node":"triage"}

id: 2
event: step
data: {"node":"retrieve"}

id: 3
event: step
data: {"node":"draft"}

id: 4
event: step
data: {"node":"check"}

id: 5
event: step
data: {"node":"reply"}

id: 6
event: step
data: {"node":"close"}

id: 7
event: delta
data: {"text":"Delivery to Córdoba typically takes 3 to 5 busin"}

id: 9
event: delta
data: {"text":"y [1]. We dispatch orders Monday to Friday [1]."}

id: 10
event: reply
data: {"text":"Delivery to Córdoba typically takes 3 to 5 business days since it is outside the warehouse's city [1]. We dispatch orders Monday to Friday [1].","sources":[{"documentId":"doc-shipping","version":3,"title":"Shipping and delivery times"}]}

id: 11
event: closed
data: {"outcome":"replied"}

What people ask first.

Do I need to code?
No. You build it, test it and launch it in the browser; your customers reach it through the API until the widget ships.
What stops it from making things up?
It can only reply citing one of your documents, and when it finds nothing, it asks you.
What if it stops and nobody is there?
The conversation waits. The shift owner gets an email and any editor can answer.
Is the demo above live?
No. It is a recorded run of the real agent on a free local model, and it shows what it replied that time.
What if I have no time to build it?
Ask us to build it. It lives in your Line account and the off switch stays yours.

Line is not open yet.

Leave your email and we will write once, when it opens.

Or ask us to build it