How to Hire a Claude Consultant: Ask About Access, Not Credentials
- Branden Bell
- 4 days ago
- 5 min read
A client gave me Owner access on their Claude Team plan. I needed it to switch on a connector, which isn't something a plain Admin can do. I built the thing, it ran, everybody was happy.
A few weeks later they asked me what I'd actually touched in there.
I answered from memory. Then I went looking for the place they could have checked without asking me, and on their plan it doesn't exist. Nobody did anything wrong. The record they wanted lives somewhere their account can't reach, and it hadn't occurred to me to say so before we started.
Now it's the first thing I bring up.
The short answer
Nobody's credentialed in this work yet, so there's nothing useful to screen on. Screen on access instead.
Three things decide it. Whose account the build runs in, what role the builder needs and when they give it back, and what record survives after they're gone. If you're still working out which surface and which plan you're even buying, that's all in my guide to using Claude in 2026.
What an owner controls
More than most buyers expect, and Anthropic publishes all of it in a table anyone can read.
Turning on a native integration, adding a custom one, or enabling a capability are Owner and Primary Owner permissions. A plain Admin can't do any of them. Admins can invite and remove members, and that's roughly where their reach stops. Modifying roles is Owner and above.
The connector model is tighter than most people assume:
Claude inherits each person's permissions from the connected service. If someone can't access a specific file, channel, or record in the source system, the connector can't reach it from Claude either.
Switching a connector on for the organization doesn't hand anybody access. Each person still authenticates individually. An Owner can hold a connected service to reading without writing across the whole organization, and users can't override that. Inside Cowork, the setting that lets somebody skip per-task approval on write-capable tools is off by default.
Anthropic engineered against the obvious risk here and did it well. The part that worries me isn't any of that.
Three settings that are already on
Cowork is on by default, and an owner has to go turn it off. On Team plans, running Cowork sessions in Anthropic's cloud is on by default too. So is the setting that allows Automatically approve mode.
None of it is hidden. It's just the state your account is in on the day you hand somebody the keys, before anyone's made a decision about any of it. If a builder doesn't walk you through those three toggles early, they haven't looked at your account yet.
The record sits outside your account
Here's the sentence that changed how I take on work:
For local sessions, Cowork stores conversation history locally on users' computers. This data is not subject to Anthropic's standard data retention policies and cannot be centrally managed or exported by admins.
Read that as a business owner instead of as a developer. If the person you hired ran their sessions on their own machine, which is the normal way to work when a build touches local files, the transcript of what they did is sitting on their laptop. Your admin can't export it. It was never in your account to export.
So you go looking for an audit log, and the roles table answers bluntly. Requesting audit logs sits under Enterprise-only controls. On the Team plan most small businesses actually buy, an Owner has nothing to request.
There's a monitoring path, and Anthropic is honest about how far it goes. Owners can stream Cowork events to a security tool through OpenTelemetry, "though it doesn't replace audit logging for compliance purposes."
The per-team controls that would let you scope a contractor down to one group are an Enterprise feature as well:
Team plans don't have access to these controls, so Cowork remains all-or-nothing.
On Team, then. Access is one switch, the record is on somebody else's computer, and the audit log is on a plan above yours. It means the record has to come out of your agreement, because it isn't coming out of your account.
Six questions about access
Ask these on the call:
Whose Claude account does this run in, and on which plan? Yours, on both counts, or you're paying to rent. I went through why that moves the number more than the hourly rate does in a separate post on what this work costs.
What role do you need, and when does it come back down? Owner access to switch on a connector is reasonable. Owner access still sitting there months after the build shipped isn't.
Directory connector, or a custom one you're writing? Anthropic's own guidance on custom connectors is that they let you connect to services "that haven't been verified by Anthropic."
Where does that custom connector run? Custom connectors get reached from Anthropic's cloud rather than from your machine, so the server has to sit on the public internet. Somebody now owns keeping that alive.
How would I know it went wrong? If you want the long version of that one, I wrote it up while working out where doing it yourself stops.
What do I get at handover, and can I read it myself? If the rules live inside their tooling, you can't change them without them.
What handover should look like
You should end up holding the account, the instructions, and the ability to change them without calling anybody.
Mine run in the client's Claude account. When the work's done they own it completely, every rule is written in plain language they can read and edit themselves, and I'm not a component in the system anymore.
Cheapest test I know, before any money moves: ask for something they built that you can run yourself. The brand kit skill is a download off my site with no email attached. The interesting part isn't the brand kit it writes, it's that the script refuses to write a file that breaks its own rules. Body text under a 7:1 contrast floor gets rejected. A positioning line over 175 characters gets rewritten rather than shrunk. A headline whose letters break apart at display size gets caught on the way out.
You can pull all of that apart without handing me anything, including your email address.
Sources
Every claim above was checked against Anthropic's own documentation in August 2026:
Want an honest read on what you'd actually be handing over? The intro call is free and runs fifteen to twenty minutes. Tell me what you want Claude to do and I'll tell you what access it genuinely needs. If we go further, consulting is $250 for the first hour, then $75 an hour, with the hours estimated up front and a fixed price agreed before any build starts.
Comments