The Marmend AI Assistant — Complete Guide

Updated July 7, 2026

The Marmend AI Assistant — Complete Guide

Marmend has a built-in AI assistant that understands plain language, works across your whole workspace, and — with your permission — can do the work, not just answer questions. This guide is the complete reference: what it is, everything it can do, how it asks before acting, how it learns your team's way of working, how it can run on its own, and how to connect it to your own AI (Claude or ChatGPT) over MCP.

In one sentence: ask (or let an event trigger it), and the assistant plans the steps, reads live data, and shows you each meaningful action for approval before it happens.

1. What it is

There are two ways to use Marmend's AI:

Inside Marmend — the AI Assistant page. Open the AI page (the sparkle icon) and type or speak a request. The assistant figures out which steps to take, gathers the data, and answers — chaining several steps in one go when needed ("find the candidates stalled in screening, then summarise each").
From your own AI — the MCP connector. Connect Claude or ChatGPT to your Marmend workspace and ask about your data (and, optionally, take actions) without leaving that assistant. See section 7.

Both use the same capabilities ("tools") and the same permission rules — what the AI can see and do always matches your role in Marmend.

Two related pages inside Marmend:

AI — where you talk to the assistant.
AI → Agent knowledge (/ai/knowledge, owner/admin) — where you set the assistant's standing objectives and review the memories it has learned. See section 5.

Availability. The AI features are enabled per workspace by your administrator or Marmend support. If you don't see the AI page, ask them to switch it on for your organization.

2. How to use it (inside Marmend)

Open the AI page.
Type or speak your request in plain language — no special syntax, no need to name tables or fields.
The assistant works through the steps and shows the result (a table, a card, a chart, or a written answer).
If a step would change data or send something out, the assistant does not do it silently. It shows an approval card — you review, then Approve, give feedback, or cancel. See section 4.

Examples of what you can just ask:

"How many candidates are in the interview stage for the Berlin backend role?"
"Show me my calendar for next week."
"Find senior Elixir engineers in our candidate base."
"Rank our candidates for the Product Designer vacancy."
"Add a note to Anna's timeline: strong on system design, follow up about salary."
"Draft a follow-up email to Mark and move him to the onsite stage." (two actions — each approved separately)

3. Everything it can do — the capability catalog

The assistant's abilities are grouped below. Read abilities run immediately. Action abilities (marked 🔒) are staged for your approval and are only available to the roles allowed to perform them.

Find, search & analyse (read — always instant)

Ability	What it does	Ask it like…
Ask about your data	Answers questions over candidates, vacancies, employees, meetings, time off, and more.	"Which vacancies have no candidates in the final stage?"
Count	Totals and headcounts.	"How many people did we hire this quarter?"
Aggregate	Average / sum / min / max of a number.	"What's the average time-to-hire for engineering roles?"
Smart people search	Finds candidates or employees by meaning, not exact keywords.	"Senior backend engineer with fintech experience."
Match candidates to a vacancy	Ranks candidates against a vacancy by fit score (best first).	"Who are the best matches for the Berlin backend role?"
Candidate context pack	Pulls everything known about one candidate for deep reasoning.	"Give me the full picture on Cara before her interview."
Meeting transcripts	Lists meetings with transcripts and answers questions about what was said.	"What did we agree with the client in Tuesday's call?"
Calendar	Shows your day or week.	"What's on my calendar Thursday?"
Free slots	Finds times when you and others are all free.	"When can I and Dana both meet for 30 minutes this week?"
World clock	Live time in one or more cities.	"What time is it in Kyiv, Berlin and San Francisco?"
Org chart	An interactive organisation chart.	"Show me the org chart for the product team."

Current people by default. Questions about employees answer about your current team — people who have left aren't counted unless you ask ("including people who left"). The same applies to archived candidates and closed vacancies.

Act on candidates 🔒

Ability	What it does	Who can	Needs approval
Move candidate stage	Moves a candidate to a different pipeline stage.	Recruiter, Owner/Admin	Yes
Add a timeline note	Writes a note on a candidate's activity feed.	Recruiter, Owner/Admin	Yes

Vacancies 🔒

Ability	What it does	Who can	Needs approval
Create a vacancy	Drafts a full job posting from a plain-language brief (saved as a draft — publishing is separate).	Recruiter, Owner/Admin	Yes
Update a vacancy	Refines or rewrites an existing vacancy.	Recruiter, Owner/Admin	Yes
Publish a vacancy	Makes the public job page live on your careers site and the job marketplace.	Recruiter, Owner/Admin	Yes

Meetings & scheduling 🔒

Ability	What it does	Who can	Needs approval
Schedule a meeting	Creates a calendar event with you as organiser and (optionally) invites attendees.	Everyone	Yes
Manage booking types	Create/edit/delete the interview & booking templates your team offers.	Owner/Admin	Yes

Communication 🔒

Ability	What it does	Who can	Needs approval
Draft an email	Prepares and validates a candidate email (subject + body) for your review — nothing is sent.	Recruiter, HR, Owner/Admin	Review only
Send an email	Sends an email to a candidate. Always staged first.	Recruiter, HR, Owner/Admin	Yes

Content 🔒

Ability	What it does	Who can	Needs approval
Create a CMS page	Drafts a page for your public site (saved as a draft).	Owner/Admin	Yes
Publish a CMS page	Renders and publishes the page live on your public site.	Owner/Admin	Yes
Create a social post	Publishes to a connected LinkedIn account or a Telegram channel.	Recruiter, Owner/Admin	Yes
Create an HR document	Creates a reusable document template (contract, NDA, certificate, …).	HR, Owner/Admin	Yes

Organisation & people 🔒 (admins)

Ability	What it does	Who can	Needs approval
Manage organisation	Create/edit/delete departments & teams, assign employees.	Owner/Admin	Yes
Employee lifecycle	Terminate, archive, or reinstate an employee (locks or restores their account).	Owner/Admin	Yes — always

The assistant's own memory

Ability	What it does
Remember / recall / refine / forget	The assistant saves and recalls lessons about how your team works (see section 5). These are internal to your workspace and are never exposed to external AI over MCP.

Note on roles. "Who can" reflects your Marmend role. If your role can't perform an action, the assistant won't offer it — and over MCP the tool simply won't be available to your connected AI.

4. How it asks before acting (approvals)

This is the assistant's most important safety rule: it never silently changes your data or sends anything outside. Reading is free; every action is staged for a human.

What you see. When the assistant wants to take an action, it shows an approval card describing exactly what it will do (and any warning, e.g. "couldn't match one attendee"). You choose:

Approve — the action runs, and the assistant continues with any remaining steps.
Give feedback — type a correction ("use the formal template", "wrong candidate") and the assistant re-plans.
Cancel — nothing happens.

Other safety properties:

Approvals survive. A pending approval is saved — if you reload the page or step away, the card is still there when you come back. If two people have the page open, whoever decides first wins; the other sees "already decided in another session" (no double action).
You always act as yourself. Meetings are created with you as organiser; the assistant can't act on someone else's behalf, and can't offboard or archive your own account.
Ambiguity stops the action. "Invite John" when there are two Johns → the assistant asks which one; nothing happens until it's clear.
Everything is recorded. Each approval (and each auto-approved action, see section 6) is logged with who decided, when, and for which action — a complete audit trail for your administrator.

5. It learns your way of working — memories & objectives

The assistant becomes your team's assistant over time. Two things shape it, both managed on the AI → Agent knowledge page (owner/admin):

Objectives — standing goals you set, applied to every request. Example: "Always reply to candidates within one business day." "Log interview notes as Strengths / Concerns / Next step."
Memories — lessons the assistant learns as it works. When you correct a staged action or give feedback, it can save the lesson so it doesn't repeat the mistake. Example: after you edit an interview note once, it remembers your format and gets it right next time.

Each memory shows its source — seeded, told by a person, or self-learned — so you can see why the assistant believes something, and edit or delete any of them. Memories are private to your workspace and are never shared with external AI over MCP.

6. Automation & autonomy

Beyond answering when asked, the assistant can work when events happen — this is what turns it from "AI you query" into "AI that works."

Event triggers

The assistant can be woken by things that happen in Marmend and act on them automatically (producing an approval for you to review):

An interview transcript is ready → it drafts an interview note and suggests a next step.
A candidate changes pipeline stage → it can follow up (e.g. draft an onboarding or rejection message).
A meeting is booked → it can prepare a briefing.

Each of these produces a pending action waiting for your approval — the work is done for you, ready to review, without you having to ask.

Scheduled runs

An assistant can run on a schedule (in your workspace's timezone) — for example a morning pipeline digest or an SLA sweep.

Trigger automations

Your existing stage automations (the "when a candidate enters this stage…" rules) gain a new option: Invoke agent. Alongside "send an email" and "move to stage", a stage can now hand off to the assistant, which decides what to do based on the candidate's context.

The autonomy slider — "do it automatically"

Once you trust the assistant with a specific low-stakes ability, you can let it skip the approval step for that ability — it runs automatically but is still recorded in the audit trail.

Eligible for autonomy (low-stakes, internal, reversible): writing timeline notes, drafting internal content, saving its own memories.
Never automatic (always needs your approval): sending emails, moving candidates between stages, scheduling meetings (external invites), and anything that publishes to the public (job posts, CMS pages, social). These "leave the building" and always stay under human control.

Autonomy is opt-in and earned: watch the assistant propose good actions, then graduate a safe ability to automatic when you're comfortable.

Availability. Automation, triggers, and autonomy are part of the AI assistant and are enabled per workspace. Ask your administrator if you'd like them switched on.

7. Connect your own AI (Claude / ChatGPT) — MCP

You can connect Marmend to Claude or ChatGPT and ask about your workspace — and optionally take actions — right from that assistant, without switching tabs. This uses a standard called MCP.

Example: "How many candidates are in the interview stage for the Berlin backend vacancy?" — answered directly in Claude, using live data from your Marmend workspace.

What your connected AI can do

Read your workspace: candidates, vacancies, employees, meetings, and smart people search.
Act (only if you grant action permissions at connection time): schedule meetings, move candidates, add timeline notes, create/update/publish vacancies, manage bookings and organisation, manage employee lifecycle, author & publish content, and send emails.
Author content with its own model. Claude/ChatGPT can write a vacancy, email, document, CMS page, or social post using their own AI, then hand it to Marmend to validate and save/publish. Marmend provides the structure and rules; the drafting is done by your AI (and isn't billed to Marmend's AI usage).

Before you start

The connector must be enabled for your organization (ask your administrator or Marmend support if you see a "not enabled" message).
Know your workspace name — the part before .marmend.com in your address (e.g. acme).
Your role decides what's available — see Permissions below.

Connect Claude (claude.ai or Claude Desktop)

Open Settings → Connectors in Claude.
Choose Add custom connector and enter: https://app.marmend.com/mcp
Claude opens a Marmend page — enter your workspace name and continue.
Sign in if needed.
Review the access screen — it shows exactly what will be readable and, separately, which actions are allowed. Press Allow access.
The Marmend tools are now available in your Claude conversations.

Connect Claude Code (terminal)

claude mcp add --transport http marmend https://app.marmend.com/mcp

Connect ChatGPT

Add https://app.marmend.com/mcp as a custom connector in ChatGPT (Settings → Connectors / custom connectors). ChatGPT's standard search/fetch is supported, so it works with both regular chats and Deep Research.

How actions stay safe over MCP

Permissions are opt-in at connection time. Grant only reading, or reading + specific actions. Without action permissions the assistant physically cannot change anything.
Your AI asks before acting. Claude/ChatGPT show a confirmation before running any action.
You act as yourself, ambiguity stops the action, skipped items are reported, and every action is logged with who, when, and which permissions — exactly like the in-app assistant.
Your assistant's private memory is never exposed over MCP.

Permissions (what each role can connect with)

Only these roles can connect an external AI, and each is limited to the abilities their role allows:

Role	Can read	Can act (if granted)
Owner / Admin	Everything	Everything (scheduling, candidates, vacancies, org, employees, content, CMS, email)
Recruiter	Candidates, vacancies, meetings	Scheduling, move/annotate candidates, author/update/publish vacancies, author & deliver content
HR	Employees, meetings	Scheduling, author & deliver content
Manager	Employees, meetings	Scheduling

Members and self-service users can't connect an external AI. Whatever your role can't do in Marmend, the connected AI can't do either.

8. Safety & privacy at a glance

Your role is the boundary. The AI — in-app or via MCP — can only see and do what your Marmend role allows. This is re-checked every time, and it fails safe (if in doubt, it does less, not more).
Nothing irreversible without a human, until you explicitly grant autonomy to a specific low-stakes ability — and even then it's logged.
Your data stays yours. Each workspace is isolated; one organisation's AI can never see another's data. The assistant's private memory and your people's sensitive details are never sent to an external AI.
Everything consequential is audited — who approved (or auto-approved) what, when.

Related guides

AI Method — the basics of asking the in-app assistant.
AI Method widgets — the interactive results (calendar, world clock, scheduling).
Agent knowledge — managing objectives and memories.
MCP connector — connecting Claude/ChatGPT in detail.
MCP external authoring — writing content with your own AI in detail.