How is this different from one connector at a time?

One connector at a time is the reconnect order after revoke or a clean start — reopen rails one by one and prove each path. Connect only what you need is the live-run habit: for this job, keep only the required connectors enabled. You use both — reconnect carefully, then run with a minimal live surface.

Why is leaving everything on a failure mode?

A specialist with the full stack live, a reconnect-all after revoke, or “I’ll just leave brokerage and mail on while I do research” means a drifted prompt can place, pay, send, or browse on rails you did not mean to use. Caps and review bound amount and the first confirm — they do not remove tools the agent can still call.

What are the common failure modes?

Leave everything on for convenience; install a specialist bot and keep the full connector stack live; reconnect-all after revoke instead of one connector at a time; leave brokerage, payer, email, or browser tools enabled for a research-only or content-only run.

Does this replace review, caps, paper, or revoke?

No. Connect only what you need limits which tools are live for the job. Review before place gates each spend or send. Caps and funding bound how much can move. Paper before live proves the pipe. Revoke before you debug closes it on incident. One connector at a time is how you reopen. You need the full stack.

When do I turn connectors back on?

When the next job truly needs that rail — and only then. Enable one required connector (or the minimum set for that desk), keep review before place on, fund only the session ceiling if money is in play, prove the path. Do not leave “maybe useful later” tools live overnight.

What is the difference between code connectors and money connectors?

Code connectors touch repos, docs, tickets, search, and files — a mistake usually costs time. Money connectors touch a brokerage, wallet, exchange, card, checkout, or anything that can sign — a mistake moves money. Email, chat, and post rails speak for you, so treat them like money for review.

Does a coding or team bot need a money connector?

No. A coding or team handoff needs code connectors only. Keep brokerage, wallet, and payer connectors off and the agent wallet empty. If a later step spends, open one money connector for that step, with spend caps written, only the session ceiling funded, and review before place on.

Can I leave code connectors on all the time?

Only for jobs that touch that code. A code connector that stays on is lower risk than a money connector, but it can still push a bad commit or close the wrong ticket. Turn on the repos and docs this job needs, and leave the rest off.

Can a code connector and a money connector be on in the same run?

Try not to. Split the job: do the code part with money off, then turn on one money connector for the step that spends. If both must be on, write your spend caps first, fund only the session amount in a dedicated agent wallet, and keep review before place on.

Where do I apply this next?

Reconnect order: one connector at a time. Specialist install: install a specialist bot. Dry prove: paper before live. Human gate: review before place. Incident: revoke before you debug. Shared computer: shared-box checklist. Hard dollar limits: agent spend caps. Browser surface: separate browser profile for the agent. Agent Lane publishes the checklist; we do not install connectors or run your agents.

Guide

Connect only what you need

Before a live agent run, enable only the MCP connectors and tools that job requires — an unused brokerage, payer, email, or browser connector still expands blast radius if the prompt drifts. Pair it with one connector at a time and review before place.

By Agent Lane Editorial · Page dated 5 October 2026

Answer first

For each live run, connect only what you need: name the job, enable only the connectors that job requires, leave unused brokerage, payer, email, browser, and post rails off, keep review before place on, and fund only if money is in the job. Leaving everything on “just in case” is how a drifted prompt spends or sends on a rail you meant to keep idle. Code connectors (repos, docs, tickets) and money connectors (brokerage, wallet, payer) are not the same risk — a coding or research run should not have a money rail live. Agent Lane publishes the checklist. We do not install connectors, place trades, hold wallets, or run your agents.

What this page is

A short operator guide for the live-run minimum surface — the habit siblings mention as “connect only what you need” when you install a specialist bot or open a spend path. Pair it with one connector at a time, paper before live, review before place, revoke before you debug, the shared-box checklist, agent spend caps, and a separate browser profile for the agent.

We are a publisher, not a connector host. Product screens change; confirm tool names on the official page before you put money in. We do not invent product menus for readers.

Why “leave everything on” is the failure mode

Full stack, one job. Research or content work with brokerage, payer, mail, and browser still live — a drifted prompt can place, pay, or send on rails you did not name.
Specialist with every connector. A marketplace bot imported for one desk inherits a spend or send path you left enabled “for later.”
Reconnect-all after revoke. Flipping every MCP back on at once reopens blast radius before you prove any path — see one connector at a time.

Caps bound how much. Review stops the first bad place or send. Connect only what you need limits which tools are callable for this job. Skip it and an unused rail is still a live rail.

What “connect only what you need” means

Job first. Name the lane in plain words — research, content, one trade prove-out, one pay — and list only the tools that lane requires.
Minimum live set. Enable those connectors; leave the rest disabled for this session even if they are installed.
Review on. Amount, destination, asset, venue, and outbound send shown before place, pay, signature, or send — review before place.
Money only if money is the job. Fund only the session ceiling when spend is in scope; otherwise keep spend connectors off and sleeves empty.
Prove, then stay minimal. Prefer paper before live on a new path. After the run, do not leave extras on overnight.

“I’ll leave brokerage and mail connected while I do research — I’m watching carefully” is not connect only what you need. Idle tools are still callable if the prompt drifts.

Code connectors vs money connectors

Short answer: a code connector reads or changes your work — repos, pull requests, docs, tickets, search, files. A money connector can move money — brokerage, wallet, exchange, card, checkout, anything that signs. A code mistake costs you time. A money mistake can cost you cash you do not get back. Keep money connectors off unless this job spends.

Connector type	Examples	Worst bad call	Default for a run
Code	GitHub, repo, docs, tickets, search, files	A messy commit or wrong comment you clean up	On only if the job touches that code
Send	Email, chat, social posts	A message goes out in your name	Off unless the job sends; review every send
Money	Brokerage, crypto wallet or exchange, card, checkout, booking	Money moves, often for good	Off unless the job spends; one at a time

A coding or team handoff needs code connectors. It does not need a wallet. If a coding bot, a team bot, or a research run has a brokerage or payer connector live “just in case,” a drifted prompt can reach money on a job that never needed it.

Code-only job: keep money connectors off and the agent wallet empty. No funding step at all.
Money job: turn on one money connector for that job. Write your agent spend caps first — a monthly budget, a per-session limit, and a daily max. Then fund only the session ceiling in a dedicated agent wallet, not your main account.
Mixed job: split it. Run the code part first with money off. Turn on the money connector only for the step that spends, with review before place on, and turn it off after.

Quick test: if this connector made one bad call right now, would you lose time or lose money? Time — it can stay on for a code job. Money — it stays off until the job spends, and then only with spend caps written and a dedicated wallet holding just the session amount.

An agent offering to book or buy something is not a reason to switch a money connector on. See a proactive ping is not permission to spend.

Where it sits in the shield stack

Dedicated agent wallet — money the bot can reach; long-term holdings stay outside.
Spend caps — session and daily maxes you write and enforce.
Paper before live — dry prove before cash on a new path.
Fund the session ceiling — empty or park excess, then put in only that ceiling when money is in the job.
Review before place — human gate on amount, destination, asset, venue, and send.
Venue — connect only what you need — for this run, only the required MCP rails are live; unused brokerage, payer, email, and browser stay off.
One connector at a time — how you reopen after revoke or a clean start.
Revoke before you debug — kill-and-isolate when something goes wrong.

Healthy run uses a minimal live surface on top of wallet, caps, funding, and review. After an incident, revoke closes the pipe; one connector at a time reopens it; this page keeps the live set small for the job.

How to connect only what you need

Name the job and the tools it needs. Write the job in plain words and list only the MCP connectors and tools that job requires — research, mail, brokerage, payer, browser, or none. Everything else stays off for this run.
Disable unused spend and send rails. Turn off brokerage, payment, email, post, and browser connectors that this job does not need. Installed is not the same as enabled for this session.
Keep review before place on. Standing rules: amount, destination, asset, venue — and outbound send — shown before any place, pay, signature, or send. No “just do it.”
Fund only if money is in the job. If the job spends: fund only the session ceiling on a dedicated path. If the job is research or content only, keep spend sleeves empty and spend connectors off.
Prove on the minimum set. Run the job with only the required connectors live. Prefer paper before live when the path is new. Confirm receipts and tool calls match the job you named.
After the run, leave extras off. Do not leave unused rails on for the next session. Reconnect later with one connector at a time if you revoked; otherwise enable only what the next job needs.

Practical checklist

Job → tool list. Write the job and the minimum connectors before you enable anything.
No “leave everything on.” Installed ≠ enabled for this session.
Specialist = still minimal. After install a specialist bot, wire only that desk’s apps — not the full stack.
After revoke: one at a time, then only what the job needs. One connector at a time, then this page for the live run.
Shared-box and browser pass. Who can sit here, which connectors are live — shared-box and separate browser profile.

Brokerage path: Shield brokerage spend. On-chain pay: Shield crypto spend and x402 buyer basics.

What Agent Lane does not do here

We do not install, enable, or revoke connectors in your Cursor or Grok account.
We do not place trades, hold wallets, or babysit your agents.
We do not publish product-menu click paths that invent screens the vendor does not ship.

Secrets in chat: see also never paste API keys in chat — vault or connector grant instead of pasting credentials into the transcript.

Related guides

Sources

Operator patterns restated from Agent Lane’s one-connector-at-a-time, install-specialist-bot, paper-before-live, review-before-place, revoke-before-debug, shared-box, spend-caps, and separate-browser-profile guides for this page dated 5 October 2026. Code vs money connector split added 5 October 2026 (evening). Tool names and product screens change — confirm on the official page before you fund a live connector. Agent Lane does not install connectors for readers.

FAQ

What does connect only what you need mean?: Connect only what you need means: before a live agent run, enable only the MCP connectors and tools that job actually requires. Leave unused brokerage, payer, email, browser, and post connectors off for that session — even if they are installed. An idle connector still expands blast radius if the prompt drifts.
How is this different from one connector at a time?: One connector at a time is the reconnect order after revoke or a clean start — reopen rails one by one and prove each path. Connect only what you need is the live-run habit: for this job, keep only the required connectors enabled. You use both — reconnect carefully, then run with a minimal live surface.
Why is leaving everything on a failure mode?: A specialist with the full stack live, a reconnect-all after revoke, or “I’ll just leave brokerage and mail on while I do research” means a drifted prompt can place, pay, send, or browse on rails you did not mean to use. Caps and review bound amount and the first confirm — they do not remove tools the agent can still call.
What are the common failure modes?: Leave everything on for convenience; install a specialist bot and keep the full connector stack live; reconnect-all after revoke instead of one connector at a time; leave brokerage, payer, email, or browser tools enabled for a research-only or content-only run.
Does this replace review, caps, paper, or revoke?: No. Connect only what you need limits which tools are live for the job. Review before place gates each spend or send. Caps and funding bound how much can move. Paper before live proves the pipe. Revoke before you debug closes it on incident. One connector at a time is how you reopen. You need the full stack.
When do I turn connectors back on?: When the next job truly needs that rail — and only then. Enable one required connector (or the minimum set for that desk), keep review before place on, fund only the session ceiling if money is in play, prove the path. Do not leave “maybe useful later” tools live overnight.
What is the difference between code connectors and money connectors?: Code connectors touch repos, docs, tickets, search, and files — a mistake usually costs time. Money connectors touch a brokerage, wallet, exchange, card, checkout, or anything that can sign — a mistake moves money. Email, chat, and post rails speak for you, so treat them like money for review.
Does a coding or team bot need a money connector?: No. A coding or team handoff needs code connectors only. Keep brokerage, wallet, and payer connectors off and the agent wallet empty. If a later step spends, open one money connector for that step, with spend caps written, only the session ceiling funded, and review before place on.
Can I leave code connectors on all the time?: Only for jobs that touch that code. A code connector that stays on is lower risk than a money connector, but it can still push a bad commit or close the wrong ticket. Turn on the repos and docs this job needs, and leave the rest off.
Can a code connector and a money connector be on in the same run?: Try not to. Split the job: do the code part with money off, then turn on one money connector for the step that spends. If both must be on, write your spend caps first, fund only the session amount in a dedicated agent wallet, and keep review before place on.
Where do I apply this next?: Reconnect order: one connector at a time. Specialist install: install a specialist bot. Dry prove: paper before live. Human gate: review before place. Incident: revoke before you debug. Shared computer: shared-box checklist. Hard dollar limits: agent spend caps. Browser surface: separate browser profile for the agent. Agent Lane publishes the checklist; we do not install connectors or run your agents.

Next steps

Before an agent touches real money, lock down a shared computer, set hard spend caps, and keep agent memory clean. Or browse all Agent Lane guides.

Get the daily roundup

The day’s useful AI agent news for small businesses, in plain language, by email. Free.