Build an AI agent without writing code
Give it a job, the documents it should read, the tools it may use, and the limits it must respect. A guided wizard walks you through all four — and shows you exactly what you built before anything goes live.
Start with the shape you need
Your first choice decides every question that follows. A scheduled agent is asked about times and recipients; a chatbot about documents and tone. Nobody answers questions that do not apply to them.
Knowledge chatbot
Answers from the material you give it — never from thin air. Cites the article it used, and hands over to a human rather than guessing.
Scheduled agent
Nobody talks to it. A clock starts it, it reads your data, and it reports what it found.
Agent with tools
Looks things up in your systems and — if you let it — changes records too. Read-only by default, per tool.
Start from nothing
Just a model and a prompt. Add knowledge and tools whenever you are ready.
Everything the builder gives you
Six things that turn a prompt into an agent you can trust in front of customers.
A wizard that explains itself
Pick what kind of agent you want and the remaining steps change to suit it. Every step says what it is, why it matters and how to answer it — before it asks you anything. Skip to the full builder at any point and come back whenever you like.
- Steps chosen by the agent you picked
- Name and brief written for you
- Nothing saved until you finish
A live map of what you actually built
Not a generic diagram. One node per knowledge base, per tool, per guardrail and per output, drawn left to right from what starts the agent to what a person ends up with. Anything you have not set up simply is not drawn — and is listed as missing.
- Read-only tools marked apart from write access
- Shows the real trigger, including a schedule
- Clicking a node opens the setting behind it
Give it knowledge in whatever form you have it
Drop in PDFs, Word files, Markdown, text, CSV, HTML or JSON. Paste an FAQ. Point at your help centre and keep it in sync. Or write it inline. Everything is indexed for meaning as well as keywords, so “can I send it back” finds an article titled “Returns”.
- Upload, paste, crawl, write, or reuse an existing base
- Meaning-based and keyword search together
- The agent can only read what you attach
Guardrails in plain English
Four questions about what the agent may see and do. From your answers the platform works out the risk and switches on the protections needed — telling you which answer caused each one. Enforced by the runtime on every step, so they hold even if the prompt is later edited.
- Live risk level as you answer
- Locked protections cannot be removed
- High-risk uses flagged before launch, not after
Tools, with permissions you choose
Your form data, the platform builder, your own API connections, and third-party MCP servers. Everything is read-only until you say otherwise, and anything that writes waits for approval unless you explicitly choose unattended operation.
- Read-only by default, per tool
- Approval gate on every write
- Every call recorded with who asked
Agents that run without being asked
Give an agent a clock instead of a chat box. It wakes on schedule, reads the data you point it at, and emails, posts to Slack, files a record — or simply logs what it found. A Run now button means you never wait to test it.
- Weekday, daily, weekly or hourly, in your timezone
- Cost ceiling per run
- Full run history with what changed
It gets better every week
An agent that never learns is an agent you eventually turn off. Mars closes the loop: every conversation is kept, the questions it could not answer become a list of what to write next, and correcting a poor answer once turns it into a permanent answer.
- See the questions it failed, ranked by how often they are asked
- Correct an answer in the log — it becomes a stored question-and-answer pair
- The gap closes itself on the next pass
And once it is live
Publishing is the beginning, not the end.
Conversation history
Every conversation kept, with the tools called, the sources cited and the confidence behind each answer.
Answer gaps
The questions it could not answer, grouped by meaning and ranked by how often they are asked. A content to-do list.
Correct it once
Fix a poor answer in the log. It becomes a question-and-answer pair and stops being wrong.
Analytics that volunteer
Volume, topics, sentiment and resolution rate — plus a written summary of what changed this week and why.
Capture leads mid-conversation
Contact details land as submissions in a form you already own, so dashboards, exports and workflows come free.
Publish anywhere
A chat widget on your site, a shareable link, or embedded in your app. Themed and branded to match.
Question-and-answer pairs
For the twenty questions that make up most of your traffic, store the exact answer. No retrieval gamble.
Access control
Decide who may view, edit, publish, chat with, and read the logs of every agent.
Safe by design, not by prompts
Guardrails are enforced outside the model, on every single step. They hold even if the prompt is ignored, edited or worked around — and the ones your answers require cannot be switched off.
It only sees what you attach
An agent pointed at your returns policy cannot read your HR documents, however it is asked.
It asks before it acts
Anything that writes shows you the exact action and waits — unless you deliberately choose unattended.
Everything is recorded
Who asked, what it called, what came back. Personal details are masked before the model ever sees them.
Your first agent in five minutes
Pick a starting point, point it at a document, and watch it answer. No credit card required.