Giving AI Hands — Lucid Solutions

Giving AI hands.

A mind that can only talk is a counselor. Give it hands and it becomes a colleague. MCP is how the hands get built — this page teaches the techniques.











The loop.

An MCP server is a small program that hands an AI a menu of safe, well-described actions over one system. Every connection, whatever the target, runs the same five beats:

  1. Advertise. The server declares its tools — names, descriptions, inputs. This text is the model's only knowledge of what's possible, so it's written with care.
  2. Ask. A person makes a request in plain language. No commands, no syntax.
  3. Choose. The model matches the request to a tool and fills in the inputs. Nobody hardcoded this decision.
  4. Execute. The server does the real work — API call, query, file write — using only the credentials it was granted.
  5. Answer. Results return to the model, which replies like a colleague, not a terminal.

Fourteen ways in.

The MCP layer is the same every time. The craft is underneath: how the server reaches the target system. Three families cover it.

Systems that welcome visitors

Modern software with a published interface. Ask politely, receive answers.

API wrapper
Auth plus endpoint calls. The default for help desks, ledgers, chat tools.an agent that finds tickets aging past their SLA and drafts the apologies.
Delegated identity
OAuth lets the server act as one specific person, with exactly their permissions, revocably.an assistant that reads your inbox and calendar — yours, not the company's.
Query language / SDK
Same courtesy, better grammar: request precisely the fields needed, nothing more.pulling three fields from a sprawling GraphQL schema in one round trip.
Webhook
Inverts the arrow — the system calls you when something happens, so nothing polls.a P1 ticket lands at 2 a.m. and triage starts before anyone wakes.
Message bus
Systems that share a bulletin board instead of talking. The agent becomes another reader or writer.consuming order events off a queue and flagging the anomalies.

Systems with no front door

Older, stubborn, or sealed software. You learn other ways in.

Command line
Slide instructions under the door: write input files, invoke the executable, collect what comes back.a card-deck tool with no API, driven by a script and CSV the agent edits, then renders.
File exchange
Watched folders and scheduled drops. Unfashionable, everywhere, and often the only path governance will bless.a nightly export lands; the agent reconciles it and files the exceptions before morning.
Direct database
Skip the application, query the data — read-only, always. Fastest when the question is about the data itself."which clients went quiet this quarter?" answered straight from the warehouse.
In-app bridge
Plant a small envoy inside the running application — a plugin on a socket — and speak to it from outside.a virtual tabletop takes requests: build tonight's journal entries from session notes.
Browser automation
A robot uses the website like a human: clicks, types, reads. Fragile — rearrange the page and it walks into walls.filing a form on a vendor portal that offers no other access.
Screen automation
The same robot for desktop programs older than their users. The true last resort.keying invoices into a green-screen ERP nobody dares replace.

Building with what's already built

The wisest technique is sometimes restraint.

Automation platform
The tool just rings a bell; a workflow someone already assembled in n8n or Make does the labor.one tool call triggers a twelve-step onboarding sequence across four systems.
Vendor servers
Software makers now ship their own MCP servers. The craft shifts from building to evaluating: which deserve write access, which need a governor.vetting a CRM vendor's server before it touches customer records.
Gateway mediation
Route every agent call through one checkpoint for identity, rate limits, and an honest logbook.the audit trail that turns a clever demo into something compliance can love.










The order of doors.

Given a new system, try the entrances in order of grace. Each step down trades elegance for reach.

The front deskAPI, SDK, delegated identity — if a published interface exists, use it.
The telephoneWebhooks and events — let the system call you.
The mailboxFiles and read-only queries — meet the data where it rests.
The envoyA bridge planted inside a sealed application.
The robotBrowser and screen automation — accepted fragility, eyes open.

The skill worth hiring is knowing the whole staircase — and never descending further than the job requires.

What a demo is made of.

The code is the smallest part of a convincing demonstration. Roughly:

Seeded data gives the agent a realistic mess to clean up. Tool descriptions are quietly load-bearing — the model chooses actions by reading them. And the scenario is chosen so the audience recognizes their own Tuesday in it. When it runs, no code appears at all. What appears is a colleague.

The systems you already run hold more reach than you know.

If you'd like help finding the doors — or building the hands — you know how to find me.