MCP & agent architecture
I design and ship Model Context Protocol servers and agent-to-agent (A2A) architectures that securely connect LLMs to your source systems — in production, not just demos.
What Model Context Protocol is and when it makes sense
Model Context Protocol (MCP) is an open standard introduced by Anthropic in 2024, defining how LLMs talk to external systems — databases, APIs, business tools — in a way that’s auditable, secure, and recoverable. I build MCP servers that work in production, not just on the demo stage.
At inFakt I lead the team that designed and shipped a production MCP server — with both read and write tools (creating invoices, clients, accounting books). Full docs: infakt.ai. The underlying problem is classic: an LLM without context from real customer data is just a pretty chatbot.
What I deliver
- Agent architecture audit — review of existing stack, identification of failure modes (auth, rate limiting, observability, fallback paths).
- MCP server implementation — from JSON-Schema spec, through auth layer (OAuth2/scoped tokens), to deployment (Docker/AWS/GCP).
- Security & compliance — GDPR review, audit logging, scoped permissions, SOC2-friendly patterns.
- Multi-provider integration — Claude / GPT / Gemini / Mistral, model-agnostic. MCP doesn’t lock you to one vendor.
- Production hardening — retry logic, idempotency, graceful degradation when backend goes down.
Common pitfalls I help avoid
- Over-broad token scopes — agents get blanket access instead of minimum required permissions.
- No audit trail — a non-starter for regulated industries.
- Synchronous calls in high-traffic contexts — MCP needs backpressure and queues.
- Single-vendor lock-in — missing the whole point of a model-agnostic standard.
Real outcome
The production MCP server at inFakt is on track to serve thousands of users via AI assistants (Claude, ChatGPT, custom systems). Wired into OCR-driven bookkeeping automation processing 300,000+ invoices monthly at 90%+ accuracy. Full case study: inFakt — MCP & OCR.
Who this fits
- SaaS companies that want to expose customer-facing AI assistant access to their data without building a chatbot.
- Enterprises with multiple source systems (ERP, CRM, data warehouse) wanting one unified LLM interface.
- Regulated companies (finance, healthcare, insurance) that need audit-friendly AI.
Stack I use
Python · TypeScript · Anthropic MCP SDK · OAuth2 · Docker · AWS Lambda · GCP Cloud Run · OpenTelemetry · LangChain (where it makes sense, not where it doesn't)