MCP Servers for Sales: Which Tools Get Write Access

InboxKit's MCP server hands an agent 74 tools behind one scope called all, and the tool that sends email has no confirm flag. We sort every tool into three piles before any can write.

Anshul
Anshul Bhatia
Founder
November 18, 2026 · 10 min read

On 17 November 2026 we counted the tools on the four MCP servers for sales that sit in our Claude Code session. InboxKit had 74. Ubersuggest had 61, Attio 56, Firecrawl 34. Some of those tools only read, and they aren't the ones to worry about. On InboxKit, by our sort, 22 other tools have no confirm flag and almost all can write, send_email among them, and all 74 sit behind a single OAuth scope whose name is "all". And Attio has 23 tools that write. Its docs say the server carries your user's permissions, and our user there is an admin, so we read those 23 as carrying an admin's reach.

Our call is to give write access by tool and never by server. Read the tool list before you connect anything, then let the tools that only read run on their own. But make every tool that buys, sends, merges, deletes or spends credits wait for a person's yes in the client.

Two situations loosen that. A server with a real read-only scope, or one where an admin can switch tools off, needs less policing on your side; we found neither on InboxKit and couldn't establish either on Attio. And a write that's free, reversible and touches only things you own can run with less ceremony. Our five test sends on 31 October 2026 went to inboxes we own, and a person still approved each one.

InboxKit is the server we've driven with writes, on 31 October and 5 November 2026. The other three we read today.

MCP servers for sales, counted tool by tool

Every count below is from our own Claude Code session's tool listing on 17 November 2026, counted by name. A client running a different version of a server may see a different number. For InboxKit there's a public check: its server card lists 74 tools, the same as our session loaded, and on 31 October a name-by-name comparison of card against live list found zero differences.

ServerTools, 17 Nov 2026Only readWrite, send or spendOur record
InboxKit743319 behind a confirm flag, 22 with none (by our sort)Driven with writes on 31 Oct and 5 Nov 2026; read again 17 Nov
Attio5632 read or render23 write, 1 only proposes a writeRead 17 Nov; no dated record of a write through this server
Ubersuggest61Not counted10 change state; generate_article spends 100 credits from the monthly poolRead 17 Nov; no dated record of a write through this server
Firecrawl34Not sorted by usMost calls spend credits; monitor tools create, run and delete schedulesRead 17 Nov; earlier work not shown to run through this server

Each of the four answered a read-only call that evening. InboxKit's get_account returned plan and wallet fields. Attio's whoami returned the signed-in member's access level, which is admin, and list-objects returned three objects. Ubersuggest's user_limits and Firecrawl's firecrawl_credit_usage returned allowances and usage counters. Nothing was written, sent or bought.

We don't add the four counts into one total. Firecrawl's 34 includes research tools that have nothing to do with sales, and a sum would read like a league table.

Two more servers sit close to our work. MoltSets, which we pay for and reviewed with measurements, offers an MCP server we haven't tested. Clay's chat MCP is separate from the API and CLI path we use for Clay, and it isn't tested by us either. Neither goes in the table.

Three piles: read, gated, open

InboxKit's tool names make the first pile easy to find. Four start with check, 16 with get, 11 with list and two with search: 4 + 16 + 11 + 2 = 33 tools that only read.

The second pile is the 19 that take a confirm flag. Seventeen list confirm as required in their schema, and two more ask for it only on some actions (manage_dns_records when deleting, manage_domain_nameservers when creating or regenerating). Five of the tools that spend money are here: buy_mailboxes, buy_prewarm_domain, register_domain, create_order and manage_wallet. InboxKit's own card says tools marked paid or destructive "require an explicit confirm flag".

And then the third pile. Twenty-two tools have no flag at all, and they can write. send_email needs only an action and a mailbox ID. update_mailbox, control_warmup and manage_mailbox_signature take no confirm parameter either, and nor do manage_sequencer, manage_webhook, manage_domain_dns_settings or create_workspace.

Bar chart of InboxKit's 74 tools in three piles: 33 only read, 19 behind a confirm flag, 22 with no flag, almost all able to write.
Our sort of InboxKit's 74 tools on 17 November 2026, from names and descriptions. A judgement, not a vendor statement.

That sort is ours, from names and descriptions, and it's a judgement. A few manage_* tools mix list, get and write actions in one tool, and manage_email_messages, for one, only lists and gets. Someone else could move a few between piles. The sort still covers all 74.

A tool that sends email from your domain, with no confirm flag, is easy to miss on a list of 74.

A confirm flag is a field the model fills in

Read the flag's own description in InboxKit's schemas: "Must be true. Ask the user for confirmation first." But that sentence is addressed to the model. The flag is a boolean in the call. A model that has already decided to buy two mailboxes sets it to true, and the server runs the call.

The MCP specification, in its 2025-06-18 text, puts the yes somewhere else. It says there SHOULD always be "a human in the loop with the ability to deny tool invocations", and that clients must treat tool annotations as untrusted unless they come from trusted servers. Attio's docs say its write operations "request user confirmation". But that's the vendor's claim, and we haven't verified how it behaves in Claude Code.

Diagram with a model and a server. A direct arrow labeled confirm: true runs the call with no person involved. A second route goes from the model through the client to a person, and only if the person says yes does the call reach the server.
Two places a yes can sit. Top: the model sets the flag and the server runs the call. Below: a person approves in the client first. Shows where a yes can sit, not what any client does.

Our own record has the approval with a person. On 31 October 2026 we drove InboxKit's server from Claude Code: reads across the account, then renames, profile pictures, removals of retired domains and five test sends to inboxes we own, with a person approving each write. For how we decide which actions get a person, see the per-action rubric for agent autonomy.

One scope means the server holds your whole login

InboxKit's server card lists one scope: "all". When we read the vendor's auth page on 31 October, it said that scope grants the same permissions the signed-in user has, and that REST keys "carry the full permissions of the account that created it". We found no read-only scope on the card or in the schemas.

Attio's docs say the server "authenticates as your Attio user" and that you have "the same permissions as your Attio user account". The page names no scopes. Our Attio login is an admin. Put those together and update-data-model is within the server's reach, along with merges and every other write on the list. That's our inference from the docs and the whoami result. But we haven't tested an Attio write through MCP to confirm it.

One InboxKit fact we can't resolve. Its own Claude-connector docs, modified 26 and 27 October 2026, say Claude can't buy mailboxes or send email, and they point at ai.inboxkit.com/mcp. The server card at mcp.inboxkit.com/mcp lists both buy_mailboxes and send_email. And we haven't established which host our session uses, so we treat the card's list as the one in front of us.

What undoes each write

Before a write runs, ask what reverses it. Several of these have no undo, and one only completes at the next renewal.

WriteServer and toolWhat undoes itWhat it costs
Merge two recordsAttio, merge-recordsNothing; both originals "can no longer be read or written"No charge stated
Remove a domainInboxKit, remove_domainsNothing stated (support, 1 Nov 2026); scheduled for deletion at next renewalNot stated
Buy mailboxesInboxKit, buy_mailboxesNo tool deletes a single cancelled mailboxOne slot per mailbox; slots used went 5 to 7, then 7 to 9 (5 Nov 2026)
Send an emailInboxKit, send_emailNothing once sentNot stated
Generate an articleUbersuggest, generate_articleCredits stay spent100 credits from the monthly pool
Scrape, search or crawlFirecrawlCredits stay spentPer call; cost not stated for most tools

But Attio's server has no delete-record tool, only deletes for comments, tasks and views. Merge is the closest thing to a delete. It makes a new record and retires both originals.

The mailbox row is where we got caught out. We ran buy_mailboxes twice on 5 November and slots used went 5 to 7, then 7 to 9, one slot per mailbox, which contradicted what support had told us in writing on 1 November.

A scrape reads a page and changes nothing you own, yet it spends credits. firecrawl_scrape reads like a read, since it only fetches a page, and it still spends credits.

How to set Claude Code permissions per server

This is the setup we recommend, not one we run today. Claude Code's permission rules can enforce the sort. Allow rules accept tool-name globs "only after a literal mcp__<server>__ prefix", so mcp__inboxkit__get_* matches InboxKit's get tools and nothing else. An unanchored allow such as mcp__* "is skipped and doesn't auto-approve anything". But in a deny rule, mcp__* does match every MCP tool. For InboxKit, four allow rules cover its 33 readers:

{
  "permissions": {
    "allow": [
      "mcp__inboxkit__check_*",
      "mcp__inboxkit__get_*",
      "mcp__inboxkit__list_*",
      "mcp__inboxkit__search_*"
    ]
  }
}

Every other InboxKit tool stays off the allow list. If your permission mode would let unlisted tools through, put them in an ask rule.

The method has a limit. A prefix only tells you how the vendor named its tools. InboxKit names its readers consistently. But Attio's tool names are hyphenated (list-records, search-records), so the globs differ and you check each match by hand. Ubersuggest names only some readers with get_ or list_ and leaves most (domain_overview, serp_analysis) unprefixed, and Firecrawl follows no read prefix, so firecrawl_scrape is a read that spends credits. Because Claude Code skips MCP rules written with parameters in parentheses, you can't allow a mixed manage_* tool for its list action only.

Claude Code's own docs say to "verify you trust each server before connecting it", and warn that servers fetching outside content can expose you to prompt injection, which we've written about for research agents.

Our own settings, read on 17 November 2026: no allow, ask or deny rule names any InboxKit, Firecrawl or Ubersuggest tool, and no prefix rule exists in any of our three settings files. The project file pre-allows four named read tools, and none of them writes. The prefix setup above is what we recommend and what the check below asks for. We don't run it today. What a given call did in a session also depends on that session's permission mode, which we haven't verified per call. The costs and the failure log of running GTM this way sit in our Claude Code for GTM piece.

The check

Run this on any server before an agent gets it, and on any vendor or agency that runs one for you. Us included.

0 of 7 checked

We can answer the last one for 31 October 2026: every write in that session had a person's yes, including the five test sends to our own inboxes.

Frequently asked questions

Which MCP servers are best for sales teams?

The ones whose tool list you've read and sorted before connecting them. We don't rank servers. On 17 November 2026 our session held InboxKit (74 tools), Ubersuggest (61), Attio (56) and Firecrawl (34). InboxKit is the only one we've driven with writes, on 31 October and 5 November 2026, and a person approved every write in the 31 October session. The other three we've read through MCP, with no dated record of a write.

Should an MCP server get write access to your CRM?

By tool, not as a whole. Attio's server has 23 tools that write, including a merge that retires both original records and a tool that updates the data model. Its docs say the server has your user's permissions, and our user is an admin. We'd let the read tools run and keep merges, updates and deletes waiting for a person's yes.

Can I limit an MCP server to read-only?

Not always at the vendor. InboxKit's server card lists one scope, "all", and Attio's docs name no scopes at all. The limit you control sits in the client. In Claude Code, allow rules with a server prefix, such as mcp__inboxkit__get_*, let read tools run without a prompt. For every other tool on that server, add an ask rule so it waits for a person.

Do MCP servers ask before they write or send?

Not by themselves. On InboxKit the confirm flag is a field in the call, and the model fills it in; its send_email tool has no flag at all. The MCP specification, in its 2025-06-18 text, says a human SHOULD be able to deny tool calls. Attio's docs say its writes request confirmation; that's the vendor's claim, and we haven't tested it in Claude Code.

How do I add an MCP server to Claude Code and see its tools?

Run claude mcp add with the http transport, a name and the server's URL. The default scope is local, which keeps the server private to you in the current project. Then open /mcp, which shows the tool count next to each connected server. Compare that count with the vendor's own server card before you allow any of its tools to run.

Supporting

  1. InboxKit, MCP server card, read 17 Nov 2026
  2. Attio, MCP overview, read 17 Nov 2026
  3. InboxKit, authentication docs, read 31 Oct 2026
  4. InboxKit, what Claude can do, read 31 Oct 2026
  5. Anthropic, Claude Code permissions, read 17 Nov 2026
  6. Anthropic, Claude Code MCP, read 17 Nov 2026
  7. Model Context Protocol, specification, tools (2025-06-18), read 17 Nov 2026
Written by
Anshul

Anshul Bhatia

Founder
IIT Kharagpur. Builds GTM systems for B2B SaaS.

Anshul builds the outbound systems behind Lead Line Partners. Clay workflows, AI enrichment, and research-first sequencing for teams that want more with less.

More posts
AI in GTMGuide · 7 min read

Claude Code for GTM: How It Works, What It Costs, What It Got Wrong

What Claude Code does in a month of GTM research, drafting and checks, what the plans cost as of November 2026, and the dated failures that show where a person has to say yes.

By Anshul Bhatia
AI in GTMPricing · 10 min read

What an AI SDR Costs in 2027: Plan Prices, Per-Meeting Math, Extras

AiSDR charges $1,500 to work 1,000 contacts on pay as you go. The same 1,000 bought as separate tools cost $157.58 as of November 2026, with labour left out of both.

By Anshul Bhatia
AI in GTMGuide · 4 min read

Ask in Trellis: Answers From Your Own Workspace, With the Sources Named

Ask opens on the account, person or call you're looking at, answers from your own Trellis workspace, and names the records behind the answer. It doesn't draft or send.

By Anshul Bhatia

Ready to engineer your GTM motion?

Tell us how your motion runs today. We'll show you what we'd engineer.

Contact us