Everything about agentor, on one page

Connect your coding agent

Your own agentor server is one address. Add it to the tool you work in, sign in with your email the first time, and say create an agent called Artur. You can do everything in the dashboard as well; the connection is what lets your coding agent teach your agent from your own files.

claude mcp add --transport http agentor https://mcp.agentor.si/mcp

{"mcpServers":{"agentor":{"url":"https://mcp.agentor.si/mcp"}}}

{"servers":{"agentor":{"type":"http","url":"https://mcp.agentor.si/mcp"}}}

{"mcpServers":{"agentor":{"serverUrl":"https://mcp.agentor.si/mcp"}}}

[mcp_servers.agentor]
url = "https://mcp.agentor.si/mcp"

Signing in: a link arrives by email and works in the browser where you asked for it; on another device, type the 6-digit code from the same email. No password. Connecting as a consultant creates your account.

Your coding agent, end to end

Connected, your own AI can create and manage your agents, read your local files into know-how, invite clients, answer open questions and watch projects. The server never reads your disk: your AI reads the files you name and sends their text. Your dashboard shows the same guide with a live connection status (and a Disconnect button) under Your coding agent.

1. Add the address to your app (above).
2. Sign in once. The first time, a browser tab opens: type your email, click the link in the mail (or type the 6-digit code), click Approve. Your AI is now signed in as you.
3. Check it works. Say agentor status or what agents do I have? — you get a YOUR AGENTS box (name, address, status, know-how items, clients, open questions), or "You have no agent yet".
4. Then say, in this order:
   1. help me create an agent — or in one go: create an agent called Artur from the technical template
   2. teach Artur from ./docs — your AI reads the folder and sends the text (it skips secrets).
   3. what is Artur missing? — the template's checklist, and the questions still open.
   4. invite ana@firma.si to Artur — Ana gets one email with everything.
   5. watch this project for Artur — one command that keeps Artur in sync by itself.
   6. show my open questions, then answer Q-3: yes, every Friday after 18:00

Lost? agentor help. Apps with a prompt menu also list /agentor:create, /agentor:start, /agentor:import, /agentor:watch, /agentor:invite, /agentor:inbox, /agentor:answer, /agentor:stats and /agentor:help (in Claude Code: /mcp__agentor__create and so on).

Something not working? Browser didn't open: in Claude Code type /mcp → agentor → Authenticate; elsewhere click Connect next to agentor. "Refused" or "expired": a sign-in link works once, for 15 minutes, in the browser that asked — start again from your app. On another computer: type the 6-digit code instead of clicking. Company proxy: your app must reach mcp.agentor.si over HTTPS. To remove it: Disconnect in the dashboard, then claude mcp remove agentor (or remove the connector in your app's settings).

Where clients come in: they never need any of this. Your invitation sends them one email with a big button to their own page; it asks which app they use and shows only those few steps.

What to say

Plain words, in any language. Your coding agent picks the right tool; help shows this list.

Your clients say even less: Artur, … with their question, any answers for me? (or ali je odgovoril?), what can you help me with? and help.

Tools

For the curious: the tools behind those phrases. Every result comes back as a short boxed summary your coding agent shows as it is.

On your server

On each agent's address, for your clients

Templates and the checklist

Start an agent from a template and it arrives with a voice, a scope, an out-of-scope line, a greeting, answering rules for your field (for example "general information, not formal advice" for legal and tax, or "ask for versions before giving commands" for technical work) and the situations it passes to you. Change any of it.

• 

  Technical consultant

  For software, devops and database consultants who keep their clients' apps running.

  25 questions your clients are likely to ask, in 9 sections

• 

  IT support / MSP

  For IT support providers who look after a company's devices, accounts, email and network.

  21 questions your clients are likely to ask, in 8 sections

• 

  Accounting & tax

  For accountants and tax advisors who keep small businesses' books, payroll and filings in order.

  20 questions your clients are likely to ask, in 8 sections

• 

  Legal advisor

  For lawyers and legal advisors who help businesses with contracts, company matters and compliance.

  20 questions your clients are likely to ask, in 8 sections

• 

  Marketing & growth

  For marketing and growth consultants who run brands, campaigns, content and analytics for clients.

  18 questions your clients are likely to ask, in 8 sections

• 

  Sales coach

  For sales coaches and consultants who train teams on pipeline, discovery, objections and closing.

  18 questions your clients are likely to ask, in 8 sections

• 

  HR & people

  For HR consultants and people partners who run hiring, onboarding, policies and reviews for companies.

  19 questions your clients are likely to ask, in 8 sections

• 

  Customer success / product support

  For product and customer success teams who help customers get started and get unstuck.

  19 questions your clients are likely to ask, in 8 sections

• 

  Business / management consultant

  For management consultants who run strategy, process and change projects with leadership teams.

  18 questions your clients are likely to ask, in 8 sections

• 

  Coach / trainer

  For coaches and trainers who run programmes, sessions and courses for people or teams.

  18 questions your clients are likely to ask, in 8 sections

• 

  Start from scratch

  A blank agent: you write the voice, scope and rules yourself.

  14 questions your clients are likely to ask, in 6 sections

Each template carries a checklist: the questions your clients are likely to ask, grouped in sections, with where that know-how usually lives in your files (README, docs, runbooks, decision records, the variable names in .env.example but never their values, the changelog, past answers). In the dashboard, open your agent's Checklist tab and answer a question right there, or say to your coding agent teach Artur from ./docs using the Technical template outline.

Readiness is the share of checklist questions your know-how covers. A question counts as covered when, asked the way a client would ask it, it finds your know-how with high confidence, using the same search your clients' questions run. Nothing is logged. Once half the checklist is covered, the dashboard suggests inviting your first client. Try it on the same page shows exactly what a client's AI would get for any question you type.

Learns by itself: watching a folder

Your agent can keep up with a project folder or git repository on your computer by itself. A small program, the watcher, runs on your computer, reads the documentation in the folders you choose and keeps your agent's know-how in step: new and changed docs become items, deleted files retire their items, and every item carries its file path, the git commit and a fingerprint (hash) of its text. The server still never reads your disk; it only receives what the watcher sends.

Install it with one line. On your agent's Learning page click Watch a project (or tell your coding agent watch this project for Artur) and run the line you get in the folder to watch. The code in it is yours alone, works once and expires after 10 minutes, so there is no separate login.

npx -y --package=https://agentor.si/dl/agentor-watch-launcher.tgz agentor-watch setup --pair K7QD-92MX

It lists exactly which files it would read and what it would make of them, asks Start watching? [Y/n], does the first sync and installs itself as a background service for your user (launchd on macOS, systemd --user on Linux, a Scheduled Task on Windows; no admin rights). Its last line tells you how to stop. It updates itself: every start checks agentor.si for a newer release signed with agentor's key and verifies it before running it.

What it reads, and never reads

• Reads: md, mdx, txt, rst, adoc and html documents; README, CHANGELOG, runbooks and decision records; OpenAPI, AsyncAPI and JSON Schema files. Code comments and docstrings only if you add --include-code-comments.
• Never reads: .env files and every other hidden file or folder, keys and certificates, credential and secret files, node_modules, vendor and build output, databases, lock files, binaries, files over 512 KB, and links that point outside the folder. In a git repository, what .gitignore ignores is left alone too.
• Documents are split at their headings, the same way every time, so a changed section updates one item. A section that still looks like a password or key is dropped on your computer before it is sent.

Review policy, per folder

Waiting items show on the Learning page ("From watch: 12 new, 3 changed") with approve, skip and approve all. They are not know-how yet, so your clients never see them. An item you edited in the dashboard is never overwritten: the file's newer version waits for you instead.

Commands

With --pull (or the pull setting), the watcher also runs git pull --ff-only for that repository before it rescans, every ten minutes. The server keeps only a display name for each folder, such as acme; the full path stays on your computer.

Suggested know-how and the weekly mail

Questions your agent could not answer well are grouped into topics over the last four weeks. Topics that keep coming back appear on the Learning page as suggested know-how to write, as the topic and how often it was asked, never a client's own words. Every Monday you get one short mail: what your agents learned that week, the topics two or more clients asked about that they couldn't answer, and the share answered from know-how this week and last. Unsubscribe from the mail itself or on the Learning page.

Writing good know-how

• One topic per item. "Database setup" and "Backups" as two items beat one long "IT" document. Long documents are split at their headings, so good headings help.
• Write the questions you get. A question-and-answer item in your own words gets a head start when a client asks the same thing, even phrased differently.
• Add pointers. "Contracts are in the shared drive under Clients / 2026" answers "where is…" questions instantly.
• Use your clients' words. If they say "the portal" and you say "the client area", add the synonym in your agent's settings.
• Test it. Say test Artur: how do we invoice? and read what comes back before a client does.
• Let the gaps teach you. The open-questions inbox is your writing list; every answer you give is saved.
• Never secrets. Keys, passwords and card numbers are refused or removed; put pointers to where they live instead.

Limits

Generous for a consulting practice, there to keep things fast and fair.

Privacy model

The server never reads your disk and runs no AI of its own. Your coding agent reads the files you name and sends their text; your clients' AI writes the answers from the passages agentor finds.

• What agentor keeps: your account email, your agents and their know-how, your clients' emails and their status, the questions clients ask with their confidence, open questions and your answers. The conversation log is deleted after 180 days.
• Who sees what: you see your own agents, clients and their questions. Each client sees only their own questions and answers, only for agents they were invited to. No consultant sees another's data.
• Secrets: API keys, tokens, private keys, card numbers, IBANs and passwords are removed from questions before they are stored or mailed, and refused as know-how.
• Know-how is data, never instructions: passages and questions reach the client's AI clearly marked as quoted material, so a sentence in a document cannot take over the conversation.
• Cookies: none on this site. Only the sign-in pages of the dashboard and the client connect pages use a sign-in cookie.

The full details are in the privacy notice and the terms.