An AI CRM with MCP for permission-aware workflows.

• Work from real CRM context

  Let assistants use the accounts, contacts, leads, deals, products, and activities already in your workspace.

• Build useful sales workflows

  Ask for reports, identify next actions, research account context, and create workflows that fit the way your team sells.

• Keep access permission-aware

  MCP actions respect the workspace access and permissions configured in Anvil.

The Model Context Protocol gives an AI client a standard way to discover and call tools. Anvil's remote MCP server describes CRM objects, readable fields, filters, and allowed actions. The assistant receives structured results instead of browser access or database credentials.

Every connection is bound to one workspace during OAuth consent. Anvil then intersects the token scopes with the stored grant, workspace switches, plan entitlements, role permissions, record visibility, and pipeline access on every call. A scope can narrow access; it cannot grant authority the member does not already have.

Most AI CRM software builds its own assistant into the product and charges for it through credits or add-ons. Anvil does not. It gives ChatGPT, Claude, and other MCP clients a controlled way into your CRM, so your team keeps the assistant it knows and pays for it once.

Use these six questions to compare any AI CRM. The answers show how Anvil handles each one.

1. Step 11

   An admin enables agents

   On Pro, a workspace admin turns on agent access under Settings → AI agents. Write access remains a separate switch and can stay off.

2. Step 22

   Add the remote endpoint

   Configure the client with https://app.anvilcrm.com/api/mcp. Anvil publishes OAuth authorization and protected-resource metadata for discovery.

3. Step 33

   Choose workspace and scopes

   Sign in, select exactly one workspace, and approve read, write, correction, or delete scopes. The client never chooses a tenant from unverified input.

4. Step 44

   Inspect before acting

   Call whoami first to see the workspace, effective scopes, CRM access, and permission matrix. Use discovery tools before sending record fields.

• Discover

  Use list_objects and describe_object to learn available object types, fields, list columns, filters, and write contracts instead of guessing payloads.

• Read and diagnose

  Search or list records, inspect activities and timelines, get pipeline summaries, and find work that needs attention within the caller's visibility.

• Act with concurrency controls

  Create records and tasks with idempotency keys. Update existing records with their expected revision so a stale assistant cannot overwrite a newer human change.

• Handle lifecycle explicitly

  Preview stage, close, reopen, conversion, ownership, and commercial actions before applying them. Dedicated tools preserve required values and lifecycle history.

• “Show the open deals I can see that have no completed activity in the last 14 days. Do not change anything.”

• “Summarize the visible timeline for Meridian Systems and list unresolved next actions with their owners.”

• “Preview moving this deal to Proposal and tell me which values or permissions are missing. Do not apply the change.”

• “Draft five follow-up tasks from this pipeline review. Show the proposed titles, owners, and due dates before creating them.”

• Live permissions

  Every call re-checks membership, roles, record access, pipeline visibility, plan entitlements, scopes, and workspace agent switches.

• Explicit high-impact confirmation

  Deletion and workspace-structure tools require the assistant to restate the change, obtain confirmation in the same turn, and send confirm: true.

• Auditable agent activity

  Tool calls record the grant, member, client, tool, outcome, duration, and touched record IDs, but not prompts, inputs, results, or field values.

• Immediate revocation

  A member can revoke their grant and an admin can revoke any workspace grant. Disabling workspace access or removing membership also blocks calls.

• Bounded requests

  Lists are paginated, request and response sizes are capped, rate limits apply, and writes use revision fences or idempotency keys.

• Human review still matters

  The server enforces authorization and data contracts; it cannot guarantee that an assistant's interpretation, summary, or proposed business decision is correct.