What are the four tools?
Four is deliberate. A small, fixed interface is easier for an agent to use correctly and easier for a company to govern than one tool per screen or per table.
| Tool | What it does | What it returns |
|---|---|---|
| crolle_search | Finds the capabilities the signed-in person is authorised to use | With a query, ranked matches with their fields; without one, the map of business areas the person may reach; with an area, every capability in it. Anything not yet available is listed apart, with the missing right and who grants it |
| crolle_describe | Returns one capability's exact versioned contract | The contract an agent must follow to call the capability; it also returns the guides and runbooks to a client that reads no resources |
| crolle_query | Runs one read | The records the read returns, within the person's reach |
| crolle_execute | Stages one change | A report of what happened to the change it staged |
Alongside the tools, Crolle publishes resources the agent reads before it acts: a guide to the product, a live what I can do guide listing every read and change this person may reach, a guide to setting up a new company, a guide to how a change happens, and a runbook for each business area that ends with the operations of that area this person can use, generated live. Professional skills, such as construction estimating, are published the same way and read when a task needs them.
How do authorisation and staging work?
Authorisation is decided before the agent sees anything. crolle_search returns only what the signed-in person is authorised to use, so the agent's map of the ERP is already the person's map. When a capability is out of reach, it is not hidden or guessed at: it is listed under not yet available with the right that is missing and who grants it, so the person knows whom to ask.
Changes are staged, one at a time. crolle_execute stages a change and reports what happened to it; it does not give the agent authority to commit. The change follows the ERP's own lifecycle, in which an authorised person commits it and the audit trail records what was proposed, who decided and on what evidence.
A workspace that is not set up yet is a state, not a fault. The records a company is set up with do not exist until an operation creates them, and an operation whose records are missing is refused by name rather than run on a guess.
Why put governance beneath the tools instead of a chatbot on top?
Because a rule enforced around an agent can be bypassed by reaching the data another way, and a rule enforced beneath the tools applies to every caller. A chatbot wired on top of several systems reads partial versions of the project and acts under none of their rules, so it cannot be trusted with an answer or allowed to act.
Crolle's rules belong to the ERP, not to the assistant. That is why the same governance applies to Crolle's own agent and to any other agent a company connects, and why an agentic ERP is a different thing from an ERP with an assistant added.
How is this different from connecting a general AI assistant to an ERP?
Connecting a general assistant to an existing ERP usually means giving it an API or a database connection under a service account, and describing the rules in a prompt. Crolle is designed the other way round.
| Question | A general assistant on an exposed ERP | Crolle's ERP over MCP |
|---|---|---|
| Whose rights does it act under? | Usually the connector's service account | The signed-in person's own rights |
| What can it see? | Whatever the connection exposes | Only the capabilities the person is authorised to use |
| What is it told about missing access? | An error, or nothing | The missing right and who grants it |
| How does it know how to call an operation? | Endpoint documentation or the prompt | An exact, versioned contract for each capability |
| What happens when it writes? | The write is applied | The change is staged for an authorised person to commit |
| Where do the rules live? | In the prompt or the wrapper | Beneath the tools, in the ERP |
| What does the audit trail show? | The service account | The person, the proposal, the decision and the evidence |
Any agent that speaks MCP can be given access to these tools, which means a company is not tied to one vendor's assistant. What it cannot do is reach around them. The controls are summarised on the security and AI governance page, and the workflows the agent is designed for are on the agent overview.
Common questions
What is Crolle's MCP interface?
It is the way AI agents read and change Crolle's construction ERP. Crolle publishes four tools over the Model Context Protocol: crolle_search, crolle_describe, crolle_query and crolle_execute. They are the only way to read or change the ERP, and every call runs under the signed-in person's rights.
Can an agent connected over MCP commit changes to the ERP?
No. crolle_execute stages one change and reports what happened to it. The change follows the ERP's own lifecycle, in which an authorised person commits it, and the audit trail records what was proposed, who decided and on what evidence.
What does an agent see when it lacks a permission?
crolle_search lists the capability apart, under not yet available, with the right that is missing and who grants it. The agent does not guess or work around the gap, and the person it works for knows exactly whom to ask.
Which AI assistants can connect to Crolle over MCP?
Any agent that speaks the Model Context Protocol can be given access to the four tools. It works under the same rights, the same staging and the same audit trail as Crolle's own agent, so a company is not tied to a single vendor's assistant to reach its own ERP.
What happens if the workspace is not set up yet?
The records a company is set up with do not exist until an operation creates them. An operation whose records are missing is refused by name, so the agent can say what has to be set up first instead of producing a result built on a guess.
Crolle is opening early access to a small group of contractors and partners. To see it on your own projects, write to info@crolle.ai.