MCP Connector Development
We connect Claude, ChatGPT or Gemini to the software your team already works in, so the AI can do real work there, with a person approving what it changes. Support desks, issue trackers, CRM, accounting, or an in-house system nobody else will touch.
What an MCP connector actually is.
MCP (Model Context Protocol) is an open standard that lets an AI assistant use your business software directly. A connector gives the AI a short, fixed list of actions it may take, like "find this invoice" or "add this customer". The AI asks, your system answers with real data. No copy-paste, and no robot clicking through screens that breaks when a button moves.
It uses your own accounts.
The connector signs in as you, to your own accounts. Cancel its access and it stops. Nothing passes through an outside company that keeps a copy.
You can see exactly what the AI is allowed to do.
Every action is written down: what it needs and what it gives back. One short list shows everything the AI can do. You do not get that with an AI clicking around a website.
It is not locked to us.
MCP is a public standard, not our own format. A connector we build for you works in Claude Desktop, Claude Code and other AI apps that support the standard. It keeps working if you stop working with us.
What we build.
- Connectors for tools that have no official AI integration and probably never will: support desks, issue trackers, niche CRMs, regional accounting software, industry systems
- Connecting an AI tool you built yourself to the software your team already works in
- Connections to your own in-house systems, so the AI can read and update your data without anyone sending it to an outside company
- Reporting connectors that can only read, when the safe answer is that the AI can look but not touch
- Connectors that can make changes, where anything that deletes or overwrites data needs your clear yes first
- Connectors for software companies that need one setup serving many customers at once
- Fixing or taking over connectors somebody else built and stopped looking after
Start small.
Not sure which workflow to connect first? Start with a review or a plan. Both are credited if you go ahead. For a whole team with several workflows, start with the audit instead: see pricing.
WORKFLOW REVIEW
S$200 (≈ US$150)
one-time
- A 45-minute session on the workflows you already run
- A one-page summary of your top three workflows to connect
- Credited toward the S$650 Automation Plan if you start it within 30 days
AUTOMATION PLAN
S$650 (≈ US$500)
one-time
- A map of the workflows we agree to cover
- The written list of AI actions, and which ones need a person’s approval
- A fixed quote and timeline for the build
- Credited in full to your first invoice
Two routes, depending on what already exists.
Set up and extend an official connector
When the vendor publishes its own MCP server (Airwallex, FlowAccount and Xero do), we configure it for your team, add the workflows and approval steps around it, and train the people who will use it.
OFFICIAL CONNECTOR SETUP
S$3,200 (≈ US$2,500)
one-time
- We configure the vendor’s own connector: Xero, Airwallex or FlowAccount
- Your workflows and approval steps added around it
- Training for the team who will use it
Build a custom connector
When there is no official connector, or it cannot do what your work needs, we build one on the public MCP standard, running on your own accounts. It starts with a S$3,200 (≈ US$2,500) proof of concept on one of your systems, credited in full.
CUSTOM CONNECTOR
From S$9,600 (≈ US$7,500)
fixed price
- After a S$3,200 (≈ US$2,500) proof of concept on one of your systems, credited in full
- Built on the public MCP standard, running on your own accounts
- Anything that changes or deletes data waits for a person
What you get either way.
- A written list of every action the AI may take: what it reads, what it can change, and what it refuses to do
- Authentication on your own accounts, with permissions limited to that list
- Approval steps: anything that changes or deletes data waits for a person
- A connector tested on your real system and installed where your team works
- Documentation, and training for the people who use it
- Defined support afterwards on a care plan, so it keeps working when models or software change
How a connector gets built.
We map your work, not the software.
The first question is never what the software can technically do. It is which decision you want to stop making by hand. Connectors planned around the software end up with forty actions nobody uses. Connectors planned around real work end up with six that get used every day.
We write down what the AI may do.
A written list of every action the connector will offer: what it reads, what it can change, and what it will refuse to do. You approve this list before any building starts, because this is where safety is decided.
We build on your real system.
We start in a practice account when the software offers one. Then we move to your live account, with every change switched off until you say so. Anything that can change your data is tested.
We install it where you work.
Claude Desktop, Claude Code, or your own app. We sit with your team the first time they use it, because good connectors usually fail in one way: nobody knows what they can do.
We keep it running.
This is the part nobody quotes for. AI models get retired, software vendors change things without notice, logins expire. On a support plan we watch what we built and fix it before you notice.
Connector catalog.
Don't see your tool? If it can already share data with other software, we'll build the connector, scoped per connector and usually shipped in 1–2 weeks. Custom connectors for your own systems are quoted at a fixed price; see pricing below.
Training for the people who use it.
Every connector includes training for your team. For broader skills, we run personal sessions and team half or full days on your own tools.
Questions we get asked.
What is MCP connector development?
MCP connector development is building a connection between one of your business systems and an AI assistant like Claude, ChatGPT or Gemini, using the open Model Context Protocol standard, so the AI can use that system directly. It covers choosing which actions the AI may take, building them, connecting safely to your own accounts, testing everything that can change data, and keeping the connector working when the software changes.
How long does a connector take to build?
A connector for well-documented software with a clear workflow is usually live in 2 to 3 weeks. Older systems, systems with little documentation, or connectors that must serve many customers at once take longer. We size those in the Automation Plan before quoting anything.
Can you build a connector for software that has no public API?
Sometimes. If the vendor offers private or partner access for other software (an API) and we can get a login for it, yes. If the only way in is the website itself, we will tell you that a robot clicking through it is fragile and usually not worth paying for. We would rather say that than sell you something that breaks in a month.
Does the AI provider train on data the connector returns?
No. Connectors call Claude, OpenAI or Google through your own accounts, under those providers business terms, which exclude training on your data. We do not store query content either.
What happens to the connector if we stop working with you?
It keeps working. The connector follows the public MCP standard and runs on your own accounts, so you can host it yourself or hand it to another vendor. There is no lock-in beyond one month notice on the recurring plan. You own the connector code we write for you; any reusable Pulsar components inside it are licensed to you, and you keep a standalone copy at handover.
Do you work with companies outside Singapore?
Yes. Pulsar Interactive Pte. Ltd. is incorporated in Singapore (UEN 202550480W) and works with clients internationally, in English.
Tell us which tool is not talking to your AI.
A free intro call, then a written plan. If a connector is the wrong answer for your case, we will say so on the call.