New Hands

The word MCP has been circling your kitchen since First Kitchen Lesson 9, where you installed one appliance — browser eyes — and learned the pantry rule: don't collect appliances. Time for the full picture, one honest sentence long: MCP is the standard plug that lets any agent connect to new abilities — browsers, databases, calendars, whatever someone's built a server for. Plug, not magic. The end.

What a brigade changes is why this matters: abilities you plug in are shared by every cook on the crew. One browser plug means your future review-specialist can see live pages, your research-hand can fetch documents, your checker can verify claims at the source. Equipping the kitchen equips everyone — which cuts both ways: clutter slows everyone too. So today runs on discipline: audit the gap, fill the gap, write the receipt.

What you'll plate today

Two new abilities your crew demonstrably lacked — chosen from evidence, installed, tested with a real task each, and receipted in the handbook so the shelf stays lean.

Ingredients

• Your kitchen-brigade project with its station chart
• Your real work from the last two weeks (the audit's raw material)
• About 30 minutes

Cook

1. Audit the gap — evidence, not catalog-browsing

The beginner move is scrolling an MCP directory going "ooh." The chef's move inverts it:

Audit this kitchen's capability gaps from EVIDENCE. Search my
journal, plans, and our recent sessions for moments where work
stalled because you couldn't do something: couldn't see a live
page, couldn't reach a service, couldn't read a format, had to
ask me to fetch/paste something a tool could have fetched.

List each stall with the evidence line, then group them: which
SINGLE ability would have cleared the most stalls? Which second?
Don't propose any installs yet.

Read the stall list — it's oddly satisfying, like finding the drawer that's been sticking. Common winners: browser eyes (if you skipped FK9), filesystem reach beyond the project, a search ability, a connector to a service you export from weekly. Your gaps are yours; that's the point of auditing instead of copying someone's list.

2. Choose like a buyer, not a collector

For the top two gaps: propose one well-known, actively maintained
MCP server each. For each tell me — in plain words:

1. What it adds, in terms of MY stalls it clears
2. What it costs: what it can access, what rides along in every
   conversation (the FK9 lesson), any account/key it needs
3. The one-line case AGAINST installing it

Then wait. I decide.

The case-against line is the discipline made visible. If the agent can't state a real cost, it hasn't thought; if you can't accept the stated cost, you've saved yourself a cluttered shelf. Decide, then:

Install the two we agreed. Walk me through any account/key steps —
keys are secrets, and secrets follow the Dinner Service law: never
in files. Then restart me guidance if needed.

(/exit, claude — new hands are noticed at shift start, same as FK9.)

3. Test each hand with a real order

Abilities you haven't exercised are rumors — fire-drill rule:

Two real tasks, one per new ability — pulled from the actual
stall list, not toy examples: [e.g. "fetch the pricing page that
stalled us Tuesday and extract the table" / "read the exports
folder outside this project and inventory it"].

For each: do it, then tell me what you did differently than you
COULD have done last week.

That last question makes the upgrade legible. "Last week I'd have asked you to paste it; today I fetched and parsed it myself" — now you know precisely what you bought.

4. Write the receipts

Add to STATIONS.md a PANTRY section: each installed ability, what
it's for (the stalls it clears), the date, and the review question
— "still earning its shelf space?" And add the standing rule to
CLAUDE.md if it's not already there from FK9: new abilities need a
real task demanding them; default answer is no.

Save point: "new hands".

When it burns

• The audit found ten gaps and everything feels urgent — ten stalls usually cluster into two abilities; that's why the audit groups them. If they genuinely don't, take the top two anyway — this lesson repeats whenever evidence accumulates. The shelf grows by proof, not by mood.
• An install wants an API key and you don't know what that is — a key is a password for programs, and it gets secret-handling: the agent walks you through where it lives (never a file that travels — Dinner Service Lesson 6 law). If a server's key setup feels sketchy or undocumented, that's the case-against winning. Skip it.
• The new ability doesn't show up — the FK9 dragon: restart the session. Still missing: "show me what MCP servers you can currently see" and debug from its answer.
• Everything got slower after installing — the exact failure FK9 warned about; every ability's manual rides along. Two shouldn't hurt; if you snuck in five, the pantry rule already knows what to say. Uninstall to the evidence line.
• Your company blocks the service an ability needs — Prep Station Lesson 9 territory: the ability audit and the case-against list is exactly what you'd show IT. Ask-first beats sneak-around, every time, still.

Order up

□ A stall list built from evidence — your kitchen's real gaps
□ Two abilities chosen with costs stated and a case-against heard
□ Both tested on real stalled tasks — upgrade made legible
□ PANTRY receipts in STATIONS.md; the default-no rule stands
□ Save point "new hands"

Next up — Lesson 04: The Specialist. The kitchen has structure and new equipment. Now it hires: your first named specialist — a subagent with one job, its own instructions, and a station on the chart.

Stuck on a step? Question box below.