When Anthropic open-sourced the Model Context Protocol (MCP) in late 2024, the immediate reaction was cautiously interested but uncertain. Eighteen months later, MCP has quietly become the connectivity layer that serious agent deployments run on. Here's why it matters and what it means for how you build.
What MCP Actually Solves
Before MCP, every agent integration was bespoke. Want your agent to query a database? Write a custom tool wrapper. Need it to call a REST API? Another wrapper. Connect to a file system? Yet another. The result was a maze of one-off integrations that were hard to maintain, impossible to share between teams, and brittle when underlying APIs changed.
MCP introduces a standard client-server model where:
- MCP servers expose resources, tools, and prompts — they're the connectors to your data and systems.
- MCP clients are the AI applications (agents, IDEs, chat interfaces) that consume those capabilities.
- The protocol layer handles transport, authentication, capability negotiation, and error handling in a standardised way.
The practical effect: an MCP server you write once for your SharePoint document library can be consumed by any MCP-compatible agent or tool without modification. Build it once, use it everywhere.
The Adoption Curve Has Hit an Inflection Point
MCP's GitHub repository passed 50,000 stars earlier this year. More tellingly, the community-maintained list of MCP servers now covers over 400 integrations — spanning major databases (PostgreSQL, MongoDB, Snowflake), enterprise SaaS (Salesforce, Jira, ServiceNow), government data sources (AusTender API, data.gov.au), and developer tools (GitHub, Linear, Confluence).
The government-sector coverage is particularly relevant to the ApiSpi audience. AusTender, the Commonwealth Grants Hub, and several state procurement portals now have maintained MCP server implementations, which means connecting a Bid & Tender Response agent to live tender data is increasingly a configuration task rather than a development project.
Security Model: What Enterprise Teams Need to Know
One of the early criticisms of MCP was that the security model was immature — specifically around credential management and what a compromised MCP server could do. Version 1.1 of the spec addressed these concerns meaningfully:
- Scoped tool permissions — clients declare exactly which tools they need at connection time; servers reject calls outside that scope.
- Audit logging is now part of the transport spec, not an afterthought. Every tool call, argument set, and response is logged with a tamper-evident chain.
- Human-in-the-loop hooks allow servers to pause execution and request explicit human approval before taking irreversible actions (writing, deleting, submitting forms).
For teams operating under ASD Essential Eight or ISO 27001 controls, these additions make MCP-based deployments defensible in a way the earlier versions weren't.
How ApiSpi Uses MCP Internally
ApiSpi's agent infrastructure has been progressively migrating to MCP servers for external integrations since early 2026. The shift has had a measurable impact on integration reliability: mean time to recovery for external API failures dropped by 70% because the MCP transport layer handles retries, backoff, and graceful degradation consistently rather than leaving it to individual agent implementations.
Current MCP integrations active in ApiSpi's production environment:
- AusTender tender feed (read-only, refreshed every 4 hours)
- SharePoint document libraries (for CV and capability statement retrieval)
- Confluence knowledge bases (for knowledge management agents)
- Jira and Linear (for enterprise architecture workflows)
- Custom internal vector stores (for RAG-based retrieval)
Building Your Own MCP Server: Where to Start
If you're connecting ApiSpi agents to internal systems, writing an MCP server is more accessible than it sounds. The TypeScript and Python SDKs are mature, well-documented, and have active communities. A basic read-only server exposing a REST API typically takes 2–4 hours to build and test.
The pattern that works best for enterprise teams: start with read-only integrations (query, search, retrieve) and introduce write capabilities only after the read path is stable and audited. This mirrors the principle of least privilege and makes security reviews significantly easier.
The Competitive Landscape
Microsoft has shipped native MCP support across the Copilot Studio product line. Google is integrating MCP into Vertex AI agent deployments. OpenAI added MCP client support to the Assistants API. The protocol has effectively won the integration wars — if you're building agents for enterprise environments, MCP is the connectivity layer you should be designing for.
The open question is governance. As MCP servers proliferate, organisations will need MCP server registries, version management policies, and supply chain security practices analogous to what they apply to npm or PyPI packages. Expect tooling in this space to mature rapidly through the rest of 2026.