# The Sashimi CRM MCP server Sashimi is a sales CRM built as an MCP server. Your AI client connects to one remote URL, you sign in and approve one workspace, and the AI gets 140 tools for leads, companies, deals, tasks and email sequences. Writes come in batches of up to 100 rows, accept an idempotency key and are rate-limited per connection. By Alfonso Trocolí, founder · Updated October 27, 2026 Canonical: https://sashimicrm.com/mcp **Sashimi MCP tool families (as of October 26, 2026)** | Family | Example tools | Writes? | | --- | --- | --- | | Schema and search | crm_schema, search_crm, get_entity, find_duplicates | No | | Leads, companies and deals | list_leads, create_lead, update_deal, merge_companies | Yes | | Bulk create and update | create_leads, create_companies, update_deals, create_tasks | Yes | | Tasks and activities | list_tasks, complete_tasks, log_activity, pin_activity | Yes | | Sequences and email templates | create_sequence, launch_sequence, list_sequence_queue, create_email_template | Yes | | Custom fields, statuses, stages and activity types | create_custom_field, set_custom_fields, create_status, create_deal_stage | Yes | | Views, reports and export | list_views, create_view, crm_report, crm_metrics, export_records | Yes (views) | | Trash | delete_leads, list_trash, restore_from_trash, purge_trash | Yes | | Members and connections | list_members, invite_member, list_connections, revoke_connection, list_mcp_activity | Yes | | Webhooks | create_webhook, list_webhooks, list_webhook_deliveries | Yes | | Prospect data (Apollo) | search_prospects, import_prospects, enrich_lead, integrations_status | Yes | | Built-in agents (off by default) | list_agents, list_agent_templates, run_agent_template | Yes | | On-demand screens (beta) | get_ui, propose_ui_patch, undo_ui_change | Yes | ## What is the Sashimi MCP server? It is the whole CRM, served as one remote MCP endpoint per workspace at `https://dashboard.sashimicrm.com/api/mcp`. Your AI client adds that URL, sends you to Sashimi to sign in, and you pick the workspace and approve. There is no app to learn, and with OAuth no key to paste into the chat. Three ways to connect: - **OAuth sign-in.** The default for Claude and ChatGPT; other remote MCP clients such as Cursor use the same URL. You approve the app by name for one workspace. - **Personal link.** `https://dashboard.sashimicrm.com/api/mcp/`, for clients without sign-in. It can be created read-only. - **Bearer header.** The same public URL with `Authorization: Bearer `, for Claude Code or your own scripts. The step-by-step guide for each client is at [Connect your AI](https://sashimicrm.com/connect). Why a CRM with no app to learn works this way is explained on the [headless CRM](https://sashimicrm.com/headless-crm) page. ## Which tools does it expose? 140 tools, grouped in the table above by family. The AI starts with `crm_schema`, which returns the workspace's statuses, stages, custom fields, activity types, sequences, email templates and members. Their keys are the ones the other tools accept, so the AI never has to guess a field name. A few things the tools do that a thin API wrapper would not: - `search_crm` turns a name into an id across leads, companies and deals. - `find_duplicates` groups records that share an email, domain or LinkedIn URL, and `merge_leads` / `merge_companies` fold them together. - `export_records` returns a CSV built from the same filters as the list tools. Built-in agents are part of the tool list but are off by default; your own AI does the reasoning. On-demand screens are beta. ## What does a connection see? Only what its member is allowed to do, in the one workspace it was approved for. The role is read live on every call. A read-only connection is never shown the write tools at all, and a member does not see admin tools such as deleting records or managing members. This is what makes one AI safe across several businesses. Each business gets its own workspace and its own connection. A connection approved for one workspace cannot read or write another. Roles and isolation are described in more detail on the [security page](https://sashimicrm.com/security), and what [one workspace per business](https://sashimicrm.com/use-cases/multiple-businesses) looks like day to day has its own page. ## How are writes built for an AI? They are designed for a model that works in batches, retries on timeouts and reads long lists. Fewer calls means fewer approvals for you and fewer turns for the model. - **Batches.** `create_leads`, `create_companies` and `create_deals` take up to 100 rows per call. `update_leads`, `update_companies`, `update_deals`, `create_tasks` and `update_tasks` work the same way. - **Per-item results.** Each row is checked on its own. Rows that fail come back with their position and the reason, and the rest are saved. - **Duplicate checks.** Bulk create checks emails against existing leads and resolves companies by name in the same call. - **Idempotency.** Every write tool accepts an optional `idempotencyKey` of 8 to 128 characters. The same key with the same arguments within 24 hours returns the first result. - **Cursor pagination.** List tools return a `nextCursor`; the list is complete only when it is null. - **Annotations.** Each tool declares `readOnlyHint`, `destructiveHint` and `openWorldHint`, so your client can ask before a delete or an outside call. Sequences follow the same care. Creating a sequence sends nothing; enrolling a record with `launch_sequence` schedules its emails. Sending goes through your own Gmail, which is beta while Google reviews Sashimi's access. ## What are the limits? Each connection can make 120 calls per minute. Within that budget, at most 20 a minute may go to tools that reach outside Sashimi, such as Apollo searches, imports and enrichment, and launching sequences. When a connection goes over, the call is refused with a message that says how many seconds to wait. The bulk tools are the way around the limit. One `update_deals` call with 40 changes counts as one call, not 40. ## Does it use credits? Only Apollo data credits, and only for prospect import and enrichment. Apollo is a third-party data provider that runs on your own Apollo key. `search_prospects` spends no credits. `import_prospects`, `enrich_lead` and `enrich_company` spend Apollo's data credits, and `integrations_status` estimates the cost before a large batch. There are no AI credits to buy: the reasoning runs on your own AI plan. What each plan includes at launch is on the [pricing](https://sashimicrm.com/pricing) page. ## How do I keep it safe? Every AI call is logged, every connection can be revoked, and nothing third parties wrote is treated as an instruction. Ask your AI to run `list_mcp_activity` to see who called which tool, on which records, and with what result. - **Untrusted text.** Notes, emails and imported fields are written by other people. Every tool result starts with a notice that its text is data, not instructions. - **Revocation.** `list_connections` and `revoke_connection` work from the AI client, and the same controls exist in the web app. - **Expiry.** Sign-in connections expire after 30 days unused and after 90 days in any case. Personal links created since 15 October 2026 expire after 90 days. - **Deletes.** Deleting records needs an admin, and deleted records stay in the trash for 30 days. Sashimi is in private beta, invitation only and free during the beta. At launch it starts at $15 a month per account (Solo, 3 workspaces). Nothing can be bought today. Using a specific client? See the pages for [Claude](https://sashimicrm.com/for/claude) and [ChatGPT](https://sashimicrm.com/for/chatgpt). The steps from adding the URL to the first call are on [how Sashimi connects](https://sashimicrm.com/how-it-works), or go back to the [home page](https://sashimicrm.com/). ## Frequently asked questions ### What is a CRM MCP server? It is a CRM that exposes its data and actions as tools over the Model Context Protocol, the open standard AI clients use to call outside services. Your AI reads and writes the CRM through those tools instead of you clicking through an app. ### Which tools does the Sashimi MCP server expose? 140 tools, covering records and search, bulk writes, tasks and activities, sequences and email templates, workspace setup, views, reports and export, trash, members and connections, webhooks and Apollo prospect data. A connection only sees the tools its member's role allows, and a read-only connection sees no write tools. ### How do batch and idempotent writes work? The bulk tools take up to 100 rows per call and report a result for each row, so one bad row does not stop the others. Every write tool also accepts an optional idempotencyKey: the same key with the same arguments within 24 hours returns the first result instead of writing again. ### What are the rate limits of the MCP server? Each connection can make 120 calls per minute. Within that, at most 20 a minute may go to tools that reach outside Sashimi, such as Apollo or launching a sequence. Over the limit, the call is refused with the number of seconds to wait. [Request beta access](https://sashimicrm.com/beta?from=mcp)