dataverse-mcp-server

MCP (Model Context Protocol) server for Microsoft Dataverse API with safe-by-default configuration. Works with any Dataverse / Dynamics 365 environment.

Tools

Data operations

Note: solution / DATAVERSE_SOLUTION_NAME only scopes list_entities (schema browsing). Data tools (query_records, get_record, create_record, …) keep full access to any table regardless of solution membership — shared tables like account or contact remain reachable.

Schema operations

Dataverse does not allow changing a column's logical name or type. To "rename" or change type: create a new column, migrate data via update_record, then delete_attribute on the old one.

Choice columns: local values or a shared Global OptionSet

A Picklist attribute takes exactly one of two fields. options defines the values inline and produces a Local OptionSet owned by that single column:

{ "logical_name": "contoso_source", "type": "Picklist", "display_name": "Source",
  "options": [ { "label": "Website", "value": 909890000 } ] }

global_option_set instead binds the column to an existing Global OptionSet by name, so several columns across several tables share one list and cannot drift apart:

{ "logical_name": "contoso_source", "type": "Picklist", "display_name": "Source",
  "global_option_set": "contoso_sourceset" }

Supplying both is rejected. That check is not cosmetic: Dataverse itself accepts the pair and then silently ignores the binding, leaving a local copy that looks bound. An unknown set name fails as Global OptionSet not found: '<name>' before anything is created — including in create_entity, which resolves names and validates every attribute before the table exists, so a rejected column cannot leave a half-built table behind.

Verify the result with get_picklist_options: a bound column reports is_global: true and the global set's own metadata_id.

Picklist option management

Picklist tools accept either entity_logical_name + attribute_logical_name (a column) or option_set_name (a Global OptionSet) — the two modes are mutually exclusive. Write operations require Customizer or System Administrator role on the connected service principal. Deleting an option does not update existing records that hold its numeric value — they are left with an orphan integer.

Telling a Global OptionSet from a local copy

get_picklist_options returns the set's identity alongside its options:

{
  "option_set": {
    "name": "fundai_source",
    "is_global": true,                                   // bound to a shared Global OptionSet
    "metadata_id": "ea6ab542-9c2e-f111-88b3-00224805d253"
  },
  "options": [ { "value": 909890000, "label": "Website" }, /* … */ ]
}

is_global is the answer to "does this column reuse an org-wide list, or does it own a private copy?" — matching values prove nothing on their own, and a column with a local set reports an auto-generated name like opportunity_prioritycode with is_global: false. To confirm which global set a column is bound to, compare its metadata_id against the one returned by get_picklist_options { option_set_name: … }.

The lookup covers Choice, Status, State and MultiSelect columns, so statecode / statuscode can be read the same way as a custom choice column.

get_entity_schema reports the same identity per column as a compact option_set summary with an option_count instead of the values themselves — read the values for a single column with get_picklist_options:

{
  "LogicalName": "fundai_source",
  "AttributeType": "Picklist",
  "option_set": { "name": "fundai_source", "is_global": true, "metadata_id": "ea6ab542-…", "option_count": 6 }
}

Actions & functions

Pass entity_set + id for a bound call (POST /<entity_set>(<id>)/Microsoft.Dynamics.CRM.<name>); omit both for an unbound call (POST /<name>). For invoke_action, parameters is the JSON request body; for invoke_function, parameters is inlined as OData function arguments. Bare operation names are namespaced automatically for bound calls — pass a fully-qualified name to override.

Examples:

// Publish a draft duplicate-detection rule.
// PublishDuplicateRule is a BOUND action on duplicaterule (returns an async job).
invoke_action({ name: "PublishDuplicateRule", entity_set: "duplicaterules", id: "<guid>" })

// Unpublish is an UNBOUND action taking DuplicateRuleId — note the asymmetry.
invoke_action({ name: "UnpublishDuplicateRule", parameters: { DuplicateRuleId: "<guid>" } })

// Qualify a lead into Account/Contact/Opportunity (bound action on lead).
invoke_action({ name: "QualifyLead", entity_set: "leads", id: "<guid>",
                parameters: { CreateAccount: true, CreateContact: true, CreateOpportunity: true, Status: 3 } })

Whether an operation is bound or unbound is defined in the Web API $metadata, not by intuition — e.g. PublishDuplicateRule is bound but UnpublishDuplicateRule is unbound. Check $metadata (look for IsBound="true" and the binding Parameter) if a call returns 404 "Resource not found for the segment".

⚠️ invoke_action can perform arbitrary mutating operations. It is currently ungated by design; capability-based access control (a safe-by-default policy gating writes/actions) is tracked separately in #45 / #46. invoke_function is read-only.

Related MCP server: xrm-mcp

Quick start (no clone)

Add to .mcp.json in your project root:

{
  "mcpServers": {
    "dataverse": {
      "command": "npx",
      "args": ["-y", "@rededis/dataverse-mcp-server"]
    }
  }
}

Create a .env file next to it with the four required variables (see Environment variables below) and restart your MCP client. The -y flag tells npx to auto-confirm the package install.

Setup

Environment variables

DATAVERSE_TENANT_ID=your-azure-tenant-id
DATAVERSE_CLIENT_ID=your-app-registration-client-id
DATAVERSE_CLIENT_SECRET=your-client-secret
DATAVERSE_RESOURCE_URL=https://your-org.crm.dynamics.com
DATAVERSE_ENTITY_PREFIX=contoso_          # optional, default prefix filter for list_entities
DATAVERSE_SOLUTION_NAME=MySolution        # optional, default solution unique name for list_entities
DATAVERSE_ALLOW_DELETE=true               # optional, enable delete operations (disabled by default)
DATAVERSE_REQUEST_TIMEOUT_MS=30000        # optional, per-request timeout in ms for Dataverse and token calls (default 30000, max 120000)
DATAVERSE_MAX_CONCURRENCY=8               # optional, requests sent to Dataverse at once (default 8, 1 to 100)
DATAVERSE_MAX_QUEUE_LENGTH=100            # optional, requests that may wait for a free slot (default 100, 0 to 10000)
DATAVERSE_MAX_QUEUE_WAIT_MS=10000         # optional, longest wait for a free slot in ms (default 10000, max 120000)
DATAVERSE_MAX_ATTEMPTS=3                  # optional, times a throttled request is sent, the first try included (default 3, 1 to 10)
DATAVERSE_MAX_RETRY_WAIT_MS=15000         # optional, longest wait on throttling per request in ms (default 15000, max 120000)

DATAVERSE_REQUEST_TIMEOUT_MS bounds how long a tool call waits for Dataverse. Dataverse itself cancels any operation after 2 minutes, hence the maximum. Schema changes such as create_entity with many columns can take longer than the 30-second default. When that happens the tool reports a timeout, but Dataverse still finishes the change, so check before retrying, or raise the value.

Service protection limits

Dataverse throttles each user with service protection limits and answers 429 Too Many Requests when one is reached. Parallel tool calls can reach them even from one machine. The server handles this in two ways, and all five DATAVERSE_MAX_* variables above are optional:

• It limits what it sends. At most DATAVERSE_MAX_CONCURRENCY requests are in flight; the rest wait in arrival order. A request is turned away with a "Dataverse is busy" error when DATAVERSE_MAX_QUEUE_LENGTH requests are already waiting, or when it has waited DATAVERSE_MAX_QUEUE_WAIT_MS without getting a slot.

• It retries a 429. The server waits for as long as the Retry-After header says, then resends, up to DATAVERSE_MAX_ATTEMPTS sends in total. While it waits, no other request is sent either, because Dataverse extends the wait for a client that keeps sending. Other requests are held back for the wait Dataverse asked for, but never longer than DATAVERSE_MAX_RETRY_WAIT_MS: a Retry-After of several minutes fails the call that received it and does not stop the rest for that long. If the waits of one request would add up to more than DATAVERSE_MAX_RETRY_WAIT_MS, the tool call fails at once with an error that says in how many seconds to retry.

With the defaults, these limits add at most 25 seconds of waiting to one request: 10 in the queue and 15 on throttling. With one request at the default 30-second timeout that makes 55 seconds, inside the 60 seconds a client built on the MCP TypeScript SDK waits for an answer unless configured otherwise. That figure assumes a 429 arrives promptly, and it does not always: in a live test, heavy requests that hit the execution-time limit were answered with 429 only after 28 to 75 seconds. The bound that holds whatever happens is 115 seconds: the 25 seconds of waiting plus three sends that each take the whole request timeout. A request that slow is cut off by DATAVERSE_REQUEST_TIMEOUT_MS first and reported as a timeout, not retried. A tool that sends several requests can take that long for each of them. If you raise the waits or DATAVERSE_REQUEST_TIMEOUT_MS, raise the tool-call timeout of your MCP client to match.

The concurrency limit alone does not prevent throttling, since Dataverse also limits combined execution time. The defaults are deliberately conservative choices of this server, not values published by Microsoft. A value that is set but out of range is reported through dataverse_setup, like a missing variable.

Azure App Registration

1. Register an app in Azure AD

2. Add API permission: Dynamics CRM > user_impersonation (or Application permissions)

3. Create a client secret

4. Grant the app a security role in Dataverse (e.g. System Administrator for full access)

Build

npm install
npm run build

npm run check is what the pre-commit hook and CI run. The pre-push hook scans the commits being pushed for environment hosts and tenant or client IDs with Betterleaks, if it is installed (brew install betterleaks, or the release binary); without it the push goes through with a warning and CI runs the scan. A git client that runs hooks with its own PATH may not find an installed Betterleaks; husky reads ~/.config/husky/init.sh, where PATH can be set.

Claude Code configuration (local build)

If you cloned the repo instead of using npx:

{
  "mcpServers": {
    "dataverse": {
      "command": "node",
      "args": ["./dist/index.js"]
    }
  }
}

Create a .env file with your credentials (see .env.example).

Safety

Destructive operations are disabled by default to prevent accidental data loss. All four delete tools are gated behind the same DATAVERSE_ALLOW_DELETE=true flag:

• delete_record — removes a row and all its data

• delete_attribute — removes a column along with ALL values across every record (no recovery short of a full environment restore)

• delete_picklist_option — removes an option from an OptionSet; records that hold the option's integer value are left with an orphan number (no label in UI, broken reports)

• delete_entity_key — drops an alternate key and its supporting unique index; any keyed-PATCH upsert flows relying on it stop working

When the flag is off, each tool registers as a stub that returns an instructional error instead of performing the delete. To enable, add DATAVERSE_ALLOW_DELETE=true to your .env file and restart the MCP server.

License

MIT