For the technical and the curious

Under the hood.

The rest of this site avoids jargon on purpose. This page does not. It uses the proper names, so anyone who knows what to look for can find it. If you just want to know what Monsel does, the other pages are shorter.

Back to the plain version

On this page

Monsel Studio knowledge + skills
Claude your team's desktop AI
Another AI tool connected with a key
Invoice agent Monsel agent
Front desk agent Monsel agent, WhatsApp
ChatGPT team members' own tool
Ops briefing agent Monsel agent, schedule

MCP

The common way AI tools connect to other software.

MCP is an open standard. A tool that speaks it can list what it offers, and an AI can call those offers. Monsel uses it in both directions.

  • Studio is an MCP server. Claude, ChatGPT and other tools that speak MCP can connect and read the same documents your agents read.
  • The Agent Orchestrator is an MCP client. Add a tool that publishes an MCP address and its actions appear in the list an agent can use.
  • What a connection can reach is set by its key, not by the tool. See the next section.

What Studio does

Keys

Every connection carries a key that opens only what you chose.

There are two kinds of connection. Both are limited to the projects you pick and both can be switched off.

  • A key is a long secret for an editor or a script. You choose which projects it opens. Keys for outside tools can be given an expiry date.
  • A sign-in connection is made by signing in once, the way you connect an app to your Google account. It uses OAuth, the usual standard for that.
  • Delete a key and it stops working straight away. A sign-in connection stops when its current access runs out. Nothing else is affected, because each connection has its own key.
  • A person can only give a connection access to projects they can see themselves.

Security in plain words

Webhooks

A message another system sends when something happens.

Your website, shop or form posts a small message to a web address Monsel gives you. The agent that listens to that address starts with the message in front of it.

  • You choose which agent listens to each address.
  • Typical uses are a new order, a submitted form or a payment that failed.
  • The agent still follows its own limits and approvals. A message from outside does not widen what it may do.

What starts an agent

Permissions

One agent can call another. Two separate checks decide whether it may.

The permission says who may ask. The approval says whether this call goes ahead.

  • An agent without permission says so plainly. If you decline an approval, the calling agent is told why.
  • Approval can also sit on the work itself: the second agent waits until a person signs off what the first one produced.
  1. Agent A asks Calls another agent
  2. Permission check Set on Agent B, default no
  3. A person approves If Agent B requires approval
  4. Agent B runs Recorded on both sides

See it as an animation

Studio and the Orchestrator

Built together, so a document can start or stop an agent.

Most platforms connect a knowledge base to an agent framework from outside. We built Studio and the Agent Orchestrator ourselves, as one project, so a change on one side reaches the other without a second integration in between.

  • A document changing project, or getting a comment, can start an agent. The trigger is the document event itself, not a workaround bolted on afterwards.
  • An agent can pause for a human decision recorded as a document in Studio. Deciding it is what lets the agent continue, or stops it there.
  • The project access you grant a person carries over to the agents that person creates, so you set it once rather than matching it up on both sides by hand.
  1. Studio A customer request is filed Added to a project in Studio
  2. Orchestrator The event starts an agent The invoice agent picks it up
  3. Orchestrator The agent drafts the invoice Reads the request, does not retype it
  4. Orchestrator The invoice is sent To the customer, nothing copied by hand
One request, filed once. Everything after it follows without anyone copying it across.

What starts an agent

A2A

A shared standard for agents on different platforms.

A2A stands for agent to agent. It lets an agent built elsewhere, or an app you are building, call one of yours, and lets yours call theirs, under rules you set.

  • A partner's agents call yours with a login you create for them. You can cancel it at any moment.
  • You choose which of your agents are visible to partners. An agent you do not expose cannot be found.
  • Your agents can call partner agents the same way, one allowed pairing at a time.
  • Calls between two workspaces you run use the same rules, with a record on both sides.
  • A product you are building can call in the same way. It does not need to be another agent: a live connection lets it use an agent you host as part of your own app, without building the orchestration, tool calls or approvals yourself.
  • That connection stays open and sends updates as the agent works, not only the final reply. Drop it and reconnect to pick up anything missed.
  • Available from the Starter plan.

Agents working together

Your own network

Reach a system inside your office or data centre without opening it to the internet.

  1. Your system A database or in-house tool
  2. One small piece Runs inside your network
  3. The agent Sees only the actions you list
Nothing of yours is opened to the internet, and no ports are opened on your firewall.

Connections in the Agent Orchestrator

Local models

Run the model on your own hardware.

An agent can use a model served by Ollama. Monsel still runs the agent: the tool calls, approvals and run records. A small piece in your network connects out to Ollama and passes the replies back.

  1. The agent loop Runs in Monsel
  2. A bridge in your network Connects out, no inbound ports
  3. Ollama Serves the model you chose
  4. The reply Relayed back to the agent
The bridge opens the connection from your side, so no inbound ports are opened on your firewall.
  • Pick the model in the agent's model list, like any other. Monsel cannot see what is installed on your machine, so you type the model name.
  • Choose a model that supports tool calls if the agent uses tools.
  • Monsel does not charge for the model, because it does not supply it. The plan price is unchanged.
  • Guardrail checks, and the rewording of a chat reply before it is sent, still call a hosted model.
  • This is separate from reaching your own systems, described under Your own network. The two can be used together.
  • Running a model this way is on the Pro plan and above. Reaching systems inside your network is on Starter and above.
  • Replies are as fast as your hardware. Test the delay before you put a customer on chat in front of it.

Marketplace

A listing is a bundle with evidence attached.

Publishing packages a folder of agents with a title, description, visibility and, if certified, the result. The plain version is on the marketplace page.

  • Visibility is public, unlisted or private. Publishing is owner-only and on the Starter plan and above. Installing is open on every plan.
  • Before install, Monsel lists what the agents need from your workspace: connections, skills, projects and any outside agents they call. Anything missing is shown, and the same list becomes the checklist after install.
  • Installed agents are created paused. Their project access is capped at what the installing user holds.
  • Certification is Monsel's own test of the agent.

Retrieval

Documents are linked when they arrive, not when you ask.

Studio prepares each document for search as it arrives. With AI enrichment switched on, it also tags it, links it to related documents and notes the people and companies it names.

  • When you ask a question, the assistant searches your documents and uses the links between them.
  • It shows which document each part of an answer came from.
  • Search never calls an AI model. The model is used to write the answer, not to find the documents.

The plain-language version

Skills

Instructions are treated like software.

An agent's behaviour is the skill that runs it. In Monsel a skill is versioned, checked when you save and tried before you publish (trying a draft is on the Pro plan and above).

Checked at save

Broken handoff paths, dangerous instructions and contradictions with what you have already taught are flagged when you click Save, not when a customer meets them.

Tried before publish

On the Pro plan and above, run a draft against past conversations or a scenario you type and read how it would answer. Pin the cases that matter and every save runs against them.

Version and restore

Every save is a version. The comparison highlights the changes that alter behaviour. One click puts a known-good version back.

Watched once it is live

On the Pro plan and above, evaluations read past runs and report which instruction caused a change in behaviour.

Written once, used everywhere

A step shared by more than one skill is written once and pulled into each of them. Change it, and every skill using it updates too, with nothing to reopen or resave.

How work gets finished

The platform checks the work, not only the AI.

A chat assistant can say it did something. An agent in Monsel has to show it. None of this makes the AI infallible. It makes failures visible and keeps a person where you put one.

  • Every tool an agent uses returns a result the platform marks as passed, needs action or failed. A failed step is not reported as success.
  • A reply that could not be delivered is not counted as done.
  • Where you guard a step with approval, the platform holds that step until a person decides.
  • Exact work is done by code. Sums and template totals are calculated, not estimated.
  • A run can wait for a person, for another agent's answer or for a set time, then carry on.
  • Each run is stored with its steps, so it can be read back and graded.

How past runs are graded

Evaluations

Agents that grade other agents' past work.

Two helper agents close the loop. Neither changes a live agent on its own.

Evaluations, which grade past runs, learning from your corrections, and trying a draft skill against past conversations are available on the Pro plan and above.

  • The Agent Evaluator reads an agent's past runs, on a schedule or on request. It reports what went wrong and which instruction caused it.
  • The Skill Health Monitor watches for the same fault coming back. It drafts a small change to one step of the instructions.
  • You approve or decline that change like any other piece of work.
  • You can add your own extra checks for an agent, and they apply to every evaluation of it.

How the loop looks in the product

Private beta

Workspaces open in batches.

Monsel is in private beta with a small number of businesses. Leave an email and we will tell you when the next batch opens. 200 free credits come with your first sign-up.

We email you when your workspace opens, and at most a handful of times before that. Every email has an unsubscribe link. Privacy.

To talk first, ask us to set agents up for your business from a written proposal, or write to hello@monsel.ai.