Chat-only (web chatbox)
- fits
- Questions, drafts, thinking out loud — no hands needed
- cost
- $0–20/mo
- complexity
- None
- upgrade when
- You catch yourself copy-pasting files into the chat window
$ equip a coding agent
A bare LLM is a brain in a jar — brilliant, and unable to touch anything. The difference between "chatbot" and "coworker" is wiring: file access, a shell, tool protocols, a remote channel, long-term memory. This guide maps every wiring shape from chat-only to multi-agent fleets, and marks the one this very site is built and operated with.
RECOMMENDED PATH
Start with a terminal agent — file access plus shell is 80% of the value. Add a skill.md for memory on day one (it's free), MCP tools when you feel the ceiling, and a phone channel only once the agent is doing real work worth interrupting your day for.
The architecture map
Every shape this can take. Most exist so you know what not to build yet.
IDE copilot (inline AI)
- fits
- You write the code yourself; AI fills in and reviews
- cost
- $0–20/mo
- complexity
- Low
- upgrade when
- You want to hand over whole tasks, not lines
Terminal agent (this site's shape)
recommended- fits
- Handing over whole tasks — build, fix, deploy, verify
- cost
- $5–50/mo
- complexity
- Low–mid
- upgrade when
- Single context can't hold the work, or you're away from the desk too often
Terminal agent + phone channel
- fits
- The agent works at home; you drive it from anywhere via chat
- cost
- $5–50/mo
- complexity
- Mid
- upgrade when
- Multiple long-running jobs start queuing behind one session
Multi-agent fleet (orchestration)
- fits
- Work too broad for one context — audits, migrations, parallel research
- cost
- $50–500+/mo
- complexity
- High
- upgrade when
- This is the frontier. If your single agent isn't saturated yet, stay put
Concepts you'll actually need
One plain sentence and one metaphor each. Click any tool module for the same treatment.
Context window
The AI's working memory — everything it can "see" right now, and it runs out.
A desk of fixed size; pile on too many papers and the oldest slide off the edge.
MCP (Model Context Protocol)
The open standard that lets any AI discover and call external tools the same way.
USB-C for AI — one plug shape, and suddenly every device fits every charger.
skill.md (portable memory)
One always-updated file holding the project's plan, stack, pitfalls and status — the next agent reads it and continues.
A shift-handover logbook; the night nurse reads it and knows exactly where things stand.
Permission boundaries
Exactly what the agent may touch — read, write, run commands, deploy — and what needs your sign-off.
Which keys you hand the live-in contractor, and which doors stay locked.
Token billing
APIs charge by text volume in and out, not per question — long files cost more than short ones.
A taxi meter — you pay for distance traveled, not for the number of rides.
Pick your budget tier
Three loadouts, one switch. Click any module for a plain-words intro with a metaphor.
The exact loadout that built this site's first version in one evening. A few dollars of API buys an agent that reads, writes, runs and deploys — you review and approve.
Take the skill.md with you
The Skill-Centric starter file for this scenario — hand it to your agent before the first prompt. It carries the plan, the stack and the known pitfalls.