Run your CRM from Claude or ChatGPT

Every action in the app is also an MCP tool. An agent gets the same permissions a person gets, and the same approval queue.

An admin approves every new account by hand. Nothing is created until then. We reply by email; no newsletter, no sequence.

  • 140+ MCP tools
  • Same scopes as the UI
  • MCP calls are not metered
app.salescrew.io/ai
Pending approvals
3 proposals · agent running
Stop all
Proposed actionActions
  • send_email
    Draft
    Re: your question about onboarding

    Thanks for the reply. Yes, imports and the first cadence are set up on the call, so there is no separate project. Would Thursday at 10 work?

    Evidence
    • Reply classified interested
    • Two open slots on the calendar
    • Not in suppression, not in cooldown
  • update_stage
  • enrich_contacts
Approve selectedReject selected

The short answer

  • SalesCrew exposes more than 140 MCP tools, covering records, inbox, outreach, calls, approvals, agent control, marketing, marketplaces, access and webhooks.
  • Every tool has a twin in the user interface with an identical effect, so turning automation off never breaks the manual path.
  • An MCP token carries the same three-axis scope model as a user account: channel, marketplace profile and area. A token can never do more than the person who issued it.
  • Any external send made through a tool goes to the approval queue by default, and every tool call is written to the audit log with the actor named.

Connect Claude in five steps

The same five steps work for ChatGPT, Cursor or a custom agent.

  1. 1

    Issue a token

    Open Settings, go to MCP tokens, and create one. Name it after the agent that will hold it, not after a person, so revoking it later is an obvious decision. The MCP twin of this screen is mcp_token_issue.

  2. 2

    Copy the server URL

    Your MCP server is an edge function on your own instance, served over HTTP. The URL belongs to your project alone, so no other customer's agent can reach it.

  3. 3

    Scope the token

    Pick the channels, the marketplace profiles and the areas this token may touch. A token for an inbox agent does not need billing or user administration. Start narrow and widen when something is genuinely blocked.

  4. 4

    Add the server to your client

    In Claude Desktop or Claude Code, add the server URL and the token to your MCP configuration and restart the client. The same two values work in any MCP client, including ChatGPT and Cursor.

  5. 5

    Run one read, then check the log

    Ask for contact_search on a name you know. When the result comes back, open the audit log and find the row. That row names the token as the actor, which is the proof the whole model rests on.

Tool families

More than 140 tools, grouped by what they touch. Example tool names shown per family.

FamilyWhat an agent can doExample tools
Contacts, companies, dealsSearch and create records, update fields, create a deal and move it between stages, query the pipeline.contact_search · deal_update_stage
InboxList threads, read one thread, draft a reply, classify a message, send an approved reply.inbox_threads · inbox_draft_reply
OutreachBuild ICPs and segments, create and pause campaigns, enrol a contact in a cadence, add a suppression entry.cadence_enroll · suppression_add
Data bankQuery the contact and company store, publish sourced rows into it from a scraper or an import.data_bank_query · data_bank_publish
LinkedIn queueAdd a profile to the human-executed queue, list what is waiting, update a status after a person sends.linkedin_queue_add · linkedin_queue_list
CallsStart an outbound call through Vapi, log a call against a contact with its outcome.call_start · call_log
Meetings and bookingsCreate a meeting, list bookings, reconcile a calendar booking against a contact record.meeting_create · booking_list
Tasks and notesCreate and close tasks, add a note to any record, list what is due.task_create · note_add
ApprovalsList what is waiting for a human, approve an item, reject one with a reason that feeds the next run.approval_list · approval_approve
Agent controlSet an agent's mode, read and write guardrails, list runs and their cost, trigger a run.agent_set_mode · agent_guardrails_set
MarketingQuery collected GA4, Search Console, Bing, ads and PostHog data, pull key figures, read channel economics.marketing_kpis · marketing_breakdown
Knowledge baseSearch the instance knowledge base, add or update an entry, list topics.kb_search · kb_upsert
MarketplacesRead the Upwork job feed, record a proposal, track an outcome, look up a marketplace client.upwork_stats · upwork_record_proposal
Billing and usageRead subscription status and the usage summary against the AI allowance.billing_status · usage_summary
Users and accessList users, set a role, grant access on a channel, profile or area.users_list · user_grant_channel
WebhooksCreate an outbound webhook, list deliveries, retry a failed one, rotate a signing secret.webhook_create · webhook_deliveries
Snapshots and fleetExport and apply an instance configuration, and register or check a deployment.snapshot_export · fleet_deployments_list

Tool count verified 2026-09-03. Snapshot export and apply are on the roadmap; the tools exist on the platform side. Copy says 140+ so the number does not go stale between releases.

What dual-control means

Dual-control is the rule that shapes this whole product: every action exists twice, once as a screen and once as a tool, and both do exactly the same thing.

That sounds like a detail. It decides whether you can trust automation. In most products the API is a subset of the interface, written later and maintained less. Some actions are screen-only, so an agent hits a wall and a human has to finish the job by hand. Other actions are API-only, so a person cannot undo what a script did. Both failures push you towards the same bad habit: giving the automation broad access so it can get past the gaps.

When the two paths match, the choice is yours every time. You can run a cadence entirely by hand, entirely from Claude, or half each. Turning every agent off with the kill switch removes nothing you need; the screens are still there and they still work. That is why the kill switch is a real control rather than a comfort blanket.

The permission model follows the same rule. A user's access is scoped on three axes: which channels they can see, which marketplace profiles they can act on, and which areas of the app they can enter. An MCP token carries those same three scopes. There is no separate permission system for automation, and no way for a token to exceed the person who created it.

What an agent can never do alone

Every instance is seeded with a policy table you can edit on the agents page. It sets the default mode for each class of action, and the defaults are deliberately cautious.

Research, enrichment, tagging, logging, scoring, summaries, internal tasks and suggested contacts run on auto. Classifying a reply, pausing a cadence when someone replies, and suppressing a contact who unsubscribed or bounced also run on auto, because the risk of doing them late is higher than the risk of doing them wrong.

Any external send — an email, a LinkedIn message, an SMS, a marketplace proposal — is set to review. Creating a deal from a positive reply and booking a meeting start in review and are worth moving to auto after a clean week. Stage changes, contact merges and deletions stay in review.

Two classes never move. Anything that spends money — model credits, marketplace connects, ad budget — and anything outside the CRM is human-only. No mode setting changes that.

On top of the modes sit guardrails: at least three days between touches per contact, a cap per cadence, daily caps per mailbox and per domain, a business-hours send window, suppression re-checked at the moment of send, and a confidence threshold that drops an agent back to review when it is unsure. Every agent ships in review mode. Nothing arrives switched on.

Four things to try on day one

Each of these is a sequence of tool calls, not a feature you have to wait for.

Triage the inbox

Ask the agent to list unread threads, classify each one, and draft replies to anything that reads as interested. The drafts arrive in the approval queue flagged as AI. You read four drafts instead of forty threads.

Build and start a segment

Describe the audience in a sentence. The agent creates the ICP, builds the segment, shows you the count before freezing, and enrols the frozen list in a cadence. Suppression applies at freeze whether a person or an agent did it.

Get a daily briefing

One call returns what came in, what is stalled, what is waiting for approval and what is due today. Run it on a schedule and read it with coffee rather than clicking through five screens.

Score a deal and set the next action

Point the agent at a deal. It reads the thread history, the company record and the pipeline stage, scores it with its reasons written out, and creates the next task. The score is a suggestion with evidence attached, not a verdict.

MCP by tier

The server itself is on every tier. What differs is whether agents run inside the app.

TierPriceMCP accessIn-product agents
Core$299/moFull server, scoped tokens, webhooks. MCP calls are not metered.None. Agents run in your own MCP client, or add the agents pack for $199/mo.
Local$499/moFull server, plus the phone and SMS tools when that pack ships.Inbox and speed-to-lead agents in review. On the roadmap.
Outbound$699/moFull server. Allowance of about 25,000 model calls a month.Campaign and LinkedIn agents. On the roadmap.
Agency$1,200/moFull server, including the marketplace and marketing tool families.Every agent. Some are on the roadmap.

MCP calls themselves are not metered on any tier. What is metered is model usage, and only when SalesCrew's own AI key does the work. Bring your own key for $99/mo and your allowance is not consumed.

Scope the token before you widen it

A token with every area enabled is a token that can update every deal and delete every contact. Merges and deletions default to review, but the safer habit is a narrow token per agent, named for the agent, revoked the day that agent stops running.

What developers and operators ask

Which MCP clients work?
Anything that speaks the Model Context Protocol. We test with Claude on desktop and in Claude Code, with ChatGPT, and with Cursor, and custom agents connect the same way. The transport is MCP over HTTP, served as an edge function on your instance.
How safe is a token?
A token is scoped on the same three axes as a user account, is stored on your side, and can be revoked from Settings or with mcp_token_revoke. Every call it makes is written to the audit log with the token named as the actor, so a token an agent holds is traceable to that agent rather than to a person.
Are there rate limits?
Tool calls are not metered, so they do not consume your AI allowance. Heavy automated use is shaped to keep one instance from affecting another.
Can an agent delete my data?
Deletions and contact merges default to review, so the action is queued and a person approves it. If you change that policy to auto you have removed the guard yourself; the audit log still records who did what. The stronger protection is scoping tokens narrowly in the first place.
Do MCP calls cost extra?
No. MCP calls are not metered on any tier. What is metered is model usage against your monthly AI allowance, and only when SalesCrew's key does the thinking. If the model runs in your own Claude client and simply calls tools here, no allowance is consumed at all.
Is there a REST API as well?
Yes. A REST API and outbound webhooks sit alongside the MCP server, so code that does not speak MCP can still read and write.
Can two agents work at the same time?
Yes. Issue a token each and scope them differently — an inbox agent that never touches billing, a research agent with read access only. Because both write to the same audit log with their own actor, you can see which agent did what without guessing.
What about Zapier, Make or n8n?
They connect through outbound webhooks and the API rather than a native app listing, and it is worth saying so plainly. Webhooks come with deliveries, retries and a rotating signing secret, so a workflow tool can subscribe to events here and act on them.

Issue a token and point Claude at it

Fourteen days. The server, the scopes, the approval queue and the audit log are all in the trial.

An admin approves every new account by hand. Nothing is created until then. We reply by email; no newsletter, no sequence.