Perth · Australia-wide · fixed price
MCP server
development.
Model Context Protocol servers, agent orchestration and tool calling — built against your real systems, not a demo. We are a Perth integration firm, so the hard part of this work, the part where a model has to reach a system that was never designed to be reached, is the part we already do.
Scoping workshop $1,500. Agent builds from $15,000. Fixed scope, published prices, and the source code is yours at handover.
All prices AUD, excluding GST. LLM API usage is billed to you by the provider, not resold by us.
The work
Four things people ask us for
Named, because you already know what you want and you should not have to translate it out of our marketing copy.
MCP servers
A server exposing your systems as tools a model can call — over stdio for a local client, or HTTP for a hosted one. Typed tool definitions, real argument validation, structured errors the model can recover from, and an auth boundary that reflects your permission model rather than the vendor's.
Agent orchestration
Multi-step workflows where the model decides the path: routing, tool selection, retries, escalation to a human, and a stopping condition that is not "hope". State kept outside the context window, so a long job survives a restart.
Retrieval that is grounded
Your documents, chunked and indexed so the answer cites the source rather than inventing it. Includes the unglamorous half: ingestion, re-indexing when the documents change, and an evaluation set so "it got better" is a measurement.
The API underneath
Usually the actual blocker. If the system has no usable API, there is nothing for a tool to call, and the agent project is really an API project wearing a hat. We build that layer — REST or GraphQL, OpenAPI documented, authenticated, tested.
Where a maintained MCP server already exists for a system you use, we will tell you to wire that one up instead of paying us to write another. That advice is free and it is usually the right answer.
The stack
What we actually work in
Not a logo wall. If it is on this list, we have shipped something in it.
Protocol & runtime
Model Context Protocol (stdio and HTTP transports), tool calling, structured outputs, streaming, Python, Node.js, TypeScript.
Orchestration
n8n, Zapier, custom queue-backed workers, webhooks and event-driven triggers, scheduled jobs, retry and dead-letter handling.
Data & retrieval
Vector indexes, document ingestion pipelines, SQL and NoSQL stores, schema design, and the migration work that usually precedes both.
Business systems
Xero, MYOB, HubSpot, Salesforce, Pipedrive, Zoho, Shopify, Google Workspace, Microsoft 365, WhatsApp Business.
Infrastructure
AWS, serverless functions, containers, CI/CD pipelines, secrets management, logging and alerting.
Measurement
GA4, Google Tag Manager, server-side tracking, Meta Conversions API, event taxonomy design — so an agent's effect is observable.
Before we quote
The questions we would ask you
Published on purpose. If a vendor has not asked you these, they will discover the answers in your production environment.
Does it have an API, or are we driving a UI?
These are different projects with different prices and different failure modes. Browser automation against a vendor's UI breaks whenever they ship, and we will say so before you pay rather than after.
What happens when it fails at 2am?
Retry, queue, alert, or stop and wait for a human — each is right in different places. The wrong answer is a silent failure that surfaces in a reconciliation six weeks later.
What is the volume and shape of the data?
Fifty documents and five million rows are not the same system. Rate limits, token cost per call and index size are all decided here, and they decide whether the running cost is $30 a month or $3,000.
What may the agent do without asking?
Reads are cheap to get wrong; writes are not. Destructive operations sit behind explicit confirmation, tools are scoped to the calling identity, and every call is logged with its arguments. A prompt instruction is not an access control.
Who maintains this in a year?
If the answer is "only the vendor", you have bought a dependency, not a system. You get source, runbook and credentials, and we would rather you never needed to call us again.
How will you know it worked?
Named before the build, not after. An evaluation set, a baseline, and a number that moves — otherwise "the AI is helping" is an opinion, and you are paying for inference to sustain it.
Price
Published, and fixed
Three entry points. If you already have the integration specified, skip the first one — we quote the build. We do not sell discovery for something you have already scoped.
Scoping workshop
$1,500
One day. We map the systems, the access paths and the constraints, and you get a written roadmap and a prioritised list. Deliberately vendor-neutral: the document is useful to a developer who is not us.
Not included: implementation, detailed technical specifications, vendor selection.
Agent build
$15,000
A retrieval-grounded agent on your own documents: conversation design, human escalation, one backend integration, admin panel and analytics. Source code included.
Not included: LLM API usage (billed to you by the provider), voice or multi-language, fine-tuning, additional integrations.
Custom API
$15,000
The layer an MCP server exposes: 15–20 REST or GraphQL endpoints over your proprietary system, OpenAPI documented, authenticated, with an automated test suite and a CI/CD pipeline.
Not included: frontend or mobile app, infrastructure costs, advanced load testing.
Send the two systems, what should happen between them, and what "done" looks like. You get a straight answer within a business day — including "this is not worth building", when that is the answer.
Straight answers
Related
Adjacent work
Automation →
One process, one week, from $990. When the job is a fixed workflow and does not need a model at all.
Systems integration →
Connecting the systems that were never designed to talk. The foundation every agent project turns out to need.
Every package →
Seventeen packages with published prices, what is included, and what is not.