Agents that act, not demos that talk.
Designing and shipping AI agents that call tools, pay for what they use and stay inside their limits. Dmitrii built Alion's agent infrastructure: an MCP server with 23 tools, keyless access, per-call payments in USDC and a personal MCP address for every person on Alion.
What an agent needsto work in production
A demo needs a prompt. A production agent needs these four layers, each designed before the first user.
MCP and function calls
Typed tools with clear descriptions, limits per caller and errors an agent can act on.
Pay per call
x402 and per-call pricing, so an agent pays for what it uses without an account or an API key.
Limits that hold
Prompt-injection filtering, data an agent may not see, and a person in the loop where money or messages move.
Every call on record
A log of every call, the cost per task and alerts when an agent loops, stalls or overspends.
Agents runningtoday
Each of these is in production and can be checked from outside.
A public MCP server
23 tools over companies, jobs and people, listed in the MCP Registry, with free keyless reads and paid calls in USDC.
An agent for a move
An AI agent that accompanies a move in the United States and an AI calculator that prices it from many parameters.
Assistants for a business
A support assistant and assistants for the marketing and sales teams of a real estate developer.
Asked beforethe first call
Which frameworks do you use?
Whatever the task needs: LangGraph, CrewAI, LlamaIndex or plain code around the model's own tool calling. The protocol (MCP) and the limits matter more than the framework.
Can our product be used by other companies' agents?
Yes, through an MCP server and an agent-readable description of what it offers and what it costs. Alion is a working example of exactly that.
How do you stop an agent from doing damage?
With limits outside the model: what each tool may do, budgets per task, filtering of untrusted text and approval where money or messages move.
Tell Dmitrii what you are building.
One message is enough: the product, the stage and what is stuck. He answers himself, usually within a day, and says plainly whether he is the right person for it.
An AI agent can do the same: check_fit with the offer, then send_message at alion.io/user/dm/mcp.
