One governed endpoint for everything your team's AI touches — Blesta, your PD modules, and the MCP servers you already use
Made and supported by Projektif Digital Ltd. Runs inside PD Modules, the panel every module installs into.
PD MCP Hub gives the AI assistants your team already uses a controlled way into Blesta. It is one secure endpoint that tools such as Claude and OpenAI Codex can connect to, so a support agent can ask an assistant about a client, an invoice or a ticket and get a real answer from your own system. What each assistant is allowed to do is decided by a connection you create for it, and nothing that matters happens without a person agreeing to it. PD MCP Hub is free.
It is built for hosts who want the benefit of AI assistants in their day-to-day work without giving any of them the keys to the business. Each assistant gets exactly the reach you grant, every call is recorded, and a serious action waits for you.
One connection per assistant, each with its own limits. Every connection has its own token, its own risk ceiling, its own areas of Blesta and the administrator it acts as. Give the assistant that reads your support queue no financial tools at all, and an instruction hidden in a ticket has nothing to reach.
A person approves anything serious. Refunds, payments, cancellations, ticket replies and anything that deletes wait in Approvals. The description is written from the actual request, not by the assistant, and shows what the assistant had just read. An approval covers only the request shown, for that connection, once, and for an hour. Rejecting tells the assistant not to retry.
Everything is on the record, including the refusals. The audit trail logs every call with the time, the connection, the tool, the outcome, how long it took and its arguments, with passwords, card details and API credentials removed. Every refusal is there too, with its reason, and the whole trail downloads as a CSV.
Two kill switches. One in the panel stops every connection immediately, even one mid-conversation. The other is a file you create in the module's folder, which stops everything over SSH when the panel itself is the problem. Delete the file and it resumes.
Safe by default. The endpoint ships switched off and answers 503 to everything until you turn it on. A read-only switch outranks every other setting, so nothing that changes data is reachable.
It grows with the modules you own. Installed PD modules add their tools on their own. A family such as DNS or SMS appears as one set of tools however many providers you have installed, and you give each connection access to the areas you choose.
PD MCP Hub is free, with every control included. It does no background work unless you attach a server, and those re-checks run on your cron, never while an assistant is waiting. It installs into PD Modules in one click and uninstalls cleanly, removing its connections, audit trail, approvals and attached servers.
When PD AI Assistant is installed too, you can switch on the bridge so its co-pilot gains the hub's tools under the connection's limits.
Yes. Every feature is included, there is nothing to upgrade to, and no part of it is held back for a paid plan — the kill switch, read-only mode, the approval queue, the full audit trail, client scoping and as many connection profiles as you need. Separate profiles are how you keep an assistant that reads tickets away from anything that moves money, so you get as many as your setup calls for. A £0 licence still ties the install to your account so updates and support work.
Exactly what the connection you gave it allows. Each connection has a risk ceiling, a set of areas it can reach, a Blesta administrator it acts as, and optionally a list of clients it may touch. The narrowest of those wins. Start read-only, watch the audit trail for a day, and widen it when you can see what it is actually doing.
Mostly the shape of the connection. Customer-written content is handed to the model marked as data rather than instructions, which helps and is not a defence on its own. The real answer is that the connection reading tickets has no financial tools at all, so an instruction saying "refund invoice 4821" has nothing to reach. Anything that does move money stops and asks a person, and the description they read is written from the actual request, with the ticket the assistant had just read shown beside it.
The kill switch on the Settings tab cuts every connection, including conversations in progress. If the panel itself is the problem, create a file called .mcp-kill in the module's directory over SSH — that works with no database, no login and no working panel.
No. Your Blesta talks directly to the assistant you connected, and directly to any third-party MCP server you attached, using your own credentials. Nothing is proxied through us and we host nothing on the path.
Anything that speaks the Model Context Protocol. Newer builds of Claude Code, Cursor and VS Code can connect to the endpoint directly over HTTPS. Claude Desktop and some others launch a local command instead, so a small bridge file is included — one file, Node 18 or newer, no dependencies, with its checksum shown in the panel so you can verify what you downloaded.
No, and that is permanent rather than a phase. This is an administrator tool with no client-area surface at all, so there is no route by which a signed-in customer could reach it.
You can connect the other tools your business already runs — Cloudflare, Stripe, GitHub, Sentry, or anything else that speaks Streamable HTTP over HTTPS — so one assistant can work across them and your Blesta together. Their tools appear under their own name, their credentials stay on your server, and nothing they offer is available until you have read the definitions and approved them.
It stops working until you look. We keep an exact record of the definitions you approved and re-check on your cron. If they change — new tools, edited descriptions, altered inputs — the server is held and shown to you as a difference, and its tools disappear from every connection until you approve the new set. That is the defence against a server that behaves for a month and then does not.
Yes, automatically. Any PD module that contributes tools makes them available here with nothing to wire up — DNS zones and records, SMS sending and history, security checks. Where several modules do the same kind of job the tools are shared, so owning one DNS module or all ten gives you one set of DNS tools, with the providers you actually own offered as choices.
It does no background work unless you attach a third-party server, and even then the re-checks run on your cron rather than while an assistant is waiting. Listing tools reads manifests and loads no module code; exactly one module is loaded when one of its tools is actually called.
An API key is all-or-nothing and has no memory. This has a ceiling per connection, a floor from the bound administrator's role, a human step in front of anything irreversible, a per-minute rate limit, and a record of every call and every refusal. It is also the difference between an assistant that can look something up and one that can quietly issue a refund.
An AI co-pilot woven into Blesta — for your staff and your customers, on the AI provider you connect
Featured
Blesta security & environment guard
Stop fake signups & order fraud before they provision
Terms and Conditions · Privacy Policy
PD Hosting is the trading name of Projektif Digital Ltd, registered in England and Wales, company number 12482555.
