CRM Account Planning Cowork Agent — MCP Server

One-click Azure deployment for the integration layer described in microsoft/ai-agent-runbooks → CRM-Account-Planning-Cowork-Agent → 2.Architecture.md.

The runbook is guidance, not code. The only component in that architecture you actually deploy to Azure is the MCP server — Copilot Cowork, the built-in skills, custom skills, and the Organization Asset Library are all M365 tenant configuration. This repo builds and deploys that MCP server, with the tool design, payload shaping, and governance switches the runbook prescribes.

The button works in any Azure tenant — it only needs Contributor on a resource group. With the default dataMode = sample it deploys and runs end to end with no CRM credentials, so you can wire it into Copilot and validate the whole path before touching Salesforce.

What gets deployed

Source is shipped as a prebuilt zip release asset and pushed to the site by the ARM deployment itself, so there is no container registry, no build pipeline in your subscription, and no secrets in the template. The deployment finishes with the code already running.

Parameters

After deployment

The deployment outputs mcpEndpoint and healthEndpoint. Check the server is up:

curl https://<siteName>.azurewebsites.net/health
# {"status":"ok","dataMode":"sample","writesEnabled":true,"protectedByApiKey":false}

Then point any MCP client at https://<siteName>.azurewebsites.net/mcp. DNS for a brand-new hostname can take a minute to propagate.

Related MCP server: Azure ML MCP Server

The tools

Five task-shaped tools, matching the runbook's "do this" table. Each returns a bounded, field-selected payload — never raw CRM records.

Design constraints honoured deliberately:

• 5 tools, not 40. Response quality degrades past ~10, and all tools of a matched plugin are injected into the prompt.

• No run_soql_query. Query composition stays server-side so the model never needs CRM schema knowledge.

• Bounded payloads. Row caps, field selection, and truncated free-text fields keep the token window intact.

• Write blast radius of one. The write tool takes exactly one record id and one field.

Connecting it to Copilot

1. Deploy, then confirm /health.

2. Register the endpoint https://<siteName>.azurewebsites.net/mcp as an MCP plugin — via Copilot Studio, or admin-deployed org-wide from the Microsoft 365 admin center (the governed path).

3. Write the plugin description carefully. Past five plugins, selection is semantic matching on the plugin description, not on individual tools.

4. Add your custom skills for account-plan structure and QBR narrative rules, and point the artifact templates at your Organization Asset Library (.potx / .dotx / .xltx).

Re-run your evaluation set after every deployment of this server: MCP plugins resolve tools dynamically at runtime, so a server change alters agent behaviour immediately with no release gate.

Identity: read this before production

The target pattern is per-user OAuth — the user authenticates to the MCP server, the server exchanges for a CRM token scoped to that user, and every tool call runs under it, so CRM record visibility follows the seller's own permissions and writes are attributable in the CRM audit trail.

The salesforce data mode in this repo implements the runbook's documented fallback: OAuth 2.0 client credentials against a connected app, i.e. a scoped service identity. Use it to get moving, and treat it as a deliberate, documented limitation:

• Restrict the plugin audience to a group whose CRM access is uniform.

• Consider allowWrites = false until per-user identity is in place.

• Validate what identity your server actually receives in the exact Copilot surface you publish to before you build on the passthrough assumption.

Move the client secret to Key Vault and reference it from app settings once you are past the pilot.

Local development

npm install
npm start          # http://localhost:3000/mcp

Environment variables: PORT, DATA_MODE, ALLOW_WRITES, MCP_API_KEY, MAX_PIPELINE_ROWS, SF_INSTANCE_URL, SF_CLIENT_ID, SF_CLIENT_SECRET, SF_API_VERSION.

Rebuild the ARM template after editing the Bicep:

az bicep build --file infra/main.bicep --outfile azuredeploy.json

Publishing a new package: push a tag (v*) and the package workflow builds the zip, attaches it to the release, and the button picks it up from releases/latest/download.

Licence

MIT. Sample data is fictional. Not an official Microsoft product.