Run your CRM from the chat window you already work in
SalesCrew ships 140+ MCP tools over HTTP. Each one has a button in the app with the same effect, the same permission scopes and the same approval queue.
- 140+ tools over MCP
- Same scopes as the UI
- Every call in the audit log
The short answer
- SalesCrew exposes 140+ MCP tools over HTTP as an edge function on your own instance, so any MCP client, whether Claude, ChatGPT, Cursor or a custom agent, can read and write the CRM.
- An MCP token carries the same three-axis scope model as a user account: channel, marketplace profile and area, so a token can never reach further than the person who issued it.
- MCP calls are not metered and do not draw down the AI allowance; what the plan meters is model calls, not tool calls.
What it does
Dual-control is the rule the product is built on: every action in the app is also a tool, and every tool has a button. Nothing is UI-only, and nothing is API-only. That is why the tool count is large — 145 tools registered as of 3 September 2026, which we round down to 140+ in copy so this page does not go stale between releases. A twenty-tool MCP server covers contacts and deals. This one covers the inbox, cadences, the data bank, approvals, agent control, marketing and the marketplace modules as well.
Transport is MCP over HTTP, served as a Supabase edge function on your instance rather than a shared gateway. Auth is a per-user token you issue in Settings or with mcp_token_issue, list with mcp_token_list and revoke with mcp_token_revoke. Each token inherits the scopes of the user who issued it across three axes: which channels, which marketplace profiles, which areas of the app. Those scopes are enforced by row-level security in Postgres, not by a check in the client. my_access returns what the current token can actually do, which is the first call worth making when something is missing.
Write safety is where the design earns its keep. Any external send routed through a tool goes to the approval queue by default, exactly as it would from the app. Stage changes, merges and deletes are review-class. Spending money is human-only. Every call, from any actor, writes to the audit log with who or what made it, so a token used badly is visible rather than deniable.
Tool families
Two examples each. The full list is returned by the server itself when a client connects.
| Family | Example tools | What an agent can do with it |
|---|---|---|
| Records | contact_search · deal_update_stage | Find people, open and move deals, read the board with pipeline_query. |
| Inbox | inbox_threads · inbox_draft_reply | Read threads, classify them, write a draft for the review queue. |
| Outreach | cadence_enroll · segment_create | Build a segment, freeze an audience, enrol contacts, pause on reply. |
| Data bank | data_bank_query · suppression_add | Query the pool, publish rows into the CRM, add to suppression. |
| LinkedIn queue | linkedin_queue_add · linkedin_queue_list | Queue a paste-ready message and read the ledger back. |
| Calls and meetings | call_start · meeting_create | Start an outbound call, log an outcome, book and update meetings. |
| Tasks and notes | task_create · note_add | Put the follow-up on someone's list and record what happened. |
| Approvals | approval_list · approval_approve | Work the review queue from the chat window, under audit. |
| Agent control | agent_set_mode · agent_guardrails_set | Change modes, read runs and cost, set the rails, hit the kill switch. |
| Marketing | marketing_kpis · marketing_niche_scoreboard | Read collected ad and organic data next to the pipeline. |
| Knowledge base | kb_search · kb_upsert | Ground a draft in your own material rather than the model's guess. |
| Marketplaces | upwork_search_list · upwork_record_proposal | Work the job feed, record proposals, track outcomes. |
| Users and access | my_access · user_grant_channel | Read what a token can do; grant scopes from the same model as the UI. |
| Webhooks | webhook_create · webhook_deliveries | Wire outbound events to Zapier, Make, n8n or your own endpoint. |
Live count: 145 registered tools as of 2026-09-03. Copy says 140+ so the number stays true between releases.
What is different here
Each of these hangs off one of the five ideas the product is built on.
Dual-control, not an API afterthought
Agent-operable by design: the manual path never breaks when automation is off, because the button and the tool are the same code path.
Read moreThe queue applies to tokens too
Human in the loop: an agent calling a send tool does not skip review. The approval queue sits under the tool, not in front of the button.
Read moreThe server runs on your instance
Your own database: the MCP endpoint is an edge function in your own Supabase project, so calls do not cross a shared gateway to reach your rows.
Read moreNot metered
MCP calls, seats, contacts and pipelines are not metered. What is metered is model calls and the usage-based channels, which is a different bill entirely.
Read moreHow you connect a client
Five minutes with a token. Same four beats as every other module.
- 1
Connect
Issue a token in Settings or with mcp_token_issue, then add the MCP URL for your instance to your client's server list. The same two values work in any MCP client.
- 2
Configure
Scope the token by channel, marketplace profile and area. Call my_access first to confirm the agent sees only what you meant it to see.
- 3
Run in review
Start with reads: contact_search, pipeline_query, inbox_threads. Leave every write agent in review so drafts queue rather than send.
- 4
Move to auto
Promote the safe classes once the audit log looks the way you expect. Revoke a token with mcp_token_revoke the moment it is not needed.
Scope the token before you hand it to an agent
my_access, and revoke it when the job is done. The audit log tells you what happened; scopes are what stop it happening.In which tier
| Tier | Price | What you get on MCP |
|---|---|---|
| Core | $299/mo | The full MCP server, scoped tokens and outbound webhooks. Agents run through MCP; add in-product agents for $199/mo. |
| Local | $499/mo | Core, plus the phone and SMS tool families. Ships with the local-services pack — on the roadmap. |
| Outbound | $699/mo | Core, plus in-product agents calling the same tools on a schedule. |
| Agency | $1,200/mo | Every module exposed, including marketplaces and marketing, across one instance per client. |
MCP calls are not metered on any tier. Seats are unlimited on every tier.
Questions
- If an agent has write access, what stops a disaster?
- Three things, in order. Token scopes limit what it can reach at all. The policy table sends external sends, stage changes, merges and deletes to review. The audit log records every call with its actor, and agent_action_revert undoes one that should not have run.
- Which clients work?
- Anything that speaks MCP. We name four because we have run them: Claude desktop and Claude Code, ChatGPT, Cursor, and custom agents built on an MCP client library.
- Is there a REST API as well?
- Yes, and outbound webhooks with deliveries and retries. A REST API and outbound webhooks sit alongside it.
- Does connecting Zapier, Make or n8n work?
- Through outbound webhooks plus the API and MCP, not through a native app listing in their directories. That is the honest mechanism, and it is worth knowing before you plan around it.
- What are the rate limits?
- Tool calls are not metered against your AI allowance. Tool calls are not metered against the AI allowance either way.
- Can two agents work at once?
- Yes. Each holds its own token with its own scopes, and each call is attributed to that token in the audit log, so you can tell which agent did what.
Issue a token and ask your CRM a question
Connect the client, run contact_search, and read the audit row it wrote.