MCP Server Runbook: Adding Internal MCP Server Integrations for Remote Platforms

This runbook provides step-by-step instructions for creating new internal MCP server integrations in Dust that connect to remote platforms (e.g., Jira, HubSpot, Salesforce, etc.).

MVP Fast Path

For a minimal new server (no OAuth, no external API yet — just the skeleton to register and test):

1. Create front/lib/api/actions/servers/{provider}/metadata.ts with a literal metadata array
2. Create front/lib/api/actions/servers/{provider}/tools/index.ts with stub handlers
3. Create front/lib/api/actions/servers/{provider}/index.ts with createServer
4. Register in constants.ts and servers/index.ts
5. Add the server to SERVER_SOURCES in bm25_tool_search_utils.test.ts
6. Add at least one BM25 query case to bm25_tool_search.test.ts

See the BM25 Tests section below for the test setup. This gives you a runnable skeleton with type-checked tool descriptions before writing any real API calls.

Quick Reference

File Structure

front/lib/api/actions/servers/{provider}/
├── metadata.ts           # Tool metadata array and server info
├── tools/index.ts        # Schema-inferred handlers and built tools
├── index.ts              # Provider-local server creation and registration
├── client.ts             # API client (optional)
└── helpers.ts            # Helper functions (optional)

Registration Files

1. front/lib/actions/mcp_internal_actions/constants.ts - Add server config with metadata: YOUR_SERVER
2. front/lib/actions/mcp_internal_actions/servers/index.ts - Import and register in switch statement

OAuth Requirements (if the platform requires OAuth)

• OAuth provider must already exist in front/lib/api/oauth/providers/{provider}.ts
• OAuth core implementation must exist in core/src/oauth/providers/{provider}.rs
• OAuth scopes must be configured for the required API access
• Server's authorization field must reference the OAuth provider

Common Gotchas

• Do not forget to add the server to AVAILABLE_INTERNAL_MCP_SERVER_NAMES array
• Server IDs must be stable and unique; never change them once deployed
• Tool stakes must be configured appropriately (never_ask, low, medium, high)
• Every internal tool must define displayLabels, toolCostCategory, and freeUsage
• Tool descriptions should start with a bare infinitive/base verb like List, Get, Search, Create, or Update
• Always implement proper error handling with Result types
• Handle OAuth token refresh automatically through the withAuth pattern

Prerequisites

OAuth Configuration (if required)

If the remote platform requires OAuth authentication:

1. Check whether an OAuth provider exists in core/src/oauth/providers/ as {provider}.rs
2. Check whether a front OAuth provider exists in front/lib/api/oauth/providers/{provider}.ts

If the OAuth provider does not exist, implement it first in core and front:

• create core/src/oauth/providers/{provider}.rs
• implement the OAuth flow: authorization URL, token exchange, refresh
• register the provider in core/src/oauth/providers/mod.rs
• create front/lib/api/oauth/providers/{provider}.ts for the front-end OAuth setup

See existing providers like hubspot.rs or jira.rs for reference implementations.

Research Phase

Before starting implementation, research the platform API:

1. API Documentation

• find the official API documentation
• identify REST endpoints vs GraphQL vs SDK usage
• note rate limits and pagination requirements

2. Authentication Method

• OAuth 2.0, preferred for user-facing integrations
• API key / bearer token, simpler but less secure
• required OAuth scopes

3. Available Operations

Document the operations you want to expose:

• read operations: list, get, search
• write operations: create, update, delete
• special operations: transitions, associations, etc.

Step-by-Step Implementation

1. Create metadata.ts

Create front/lib/api/actions/servers/{provider}/metadata.ts:

import type { ServerMetadata } from "@app/lib/actions/mcp_internal_actions/tool_definition";
import { z } from "zod";

export const YOUR_PROVIDER_SERVER_NAME = "your_provider" as const;

export const YOUR_PROVIDER_TOOLS_METADATA = [
  {
    name: "list_items",
    description: "List all items accessible to the user.",
    schema: {
      pageToken: z.string().optional().describe("Page token for pagination."),
      maxResults: z.number().optional().describe("Maximum results to return."),
    },
    stake: "never_ask",
    toolCostCategory: "advanced",
    freeUsage: false,
    displayLabels: {
      running: "Listing Items",
      done: "List items",
    },
  },
  {
    name: "get_item",
    description: "Get a single item by ID.",
    schema: {
      itemId: z.string().describe("The ID of the item to retrieve."),
    },
    stake: "never_ask",
    toolCostCategory: "advanced",
    freeUsage: false,
    displayLabels: {
      running: "Retrieving item",
      done: "Retrieve item",
    },
  },
  {
    name: "create_item",
    description: "Create a new item.",
    schema: {
      name: z.string().describe("Name of the item."),
      description: z.string().optional().describe("Description of the item."),
    },
    stake: "low",
    toolCostCategory: "advanced",
    freeUsage: false,
    displayLabels: {
      running: "Creating item",
      done: "Create item",
    },
  },
] as const;

export const YOUR_PROVIDER_SERVER = {
  serverInfo: {
    name: YOUR_PROVIDER_SERVER_NAME,
    version: "1.0.0",
    description: "Short description of what this integration does.",
    authorization: {
      provider: "your_provider",
      supported_use_cases: ["personal_actions", "platform_actions"],
    },
    icon: "YourProviderLogo",
    documentationUrl: "https://docs.dust.tt/docs/your-provider",
    instructions: null,
  },
  tools: YOUR_PROVIDER_TOOLS_METADATA,
} as const satisfies ServerMetadata;

Key points:

• Use snake_case for the tool names
• Internal tools must include displayLabels; unlike remote MCP tools, these labels are required
• Set toolCostCategory and freeUsage deliberately, following a comparable existing server
• tool descriptions start with a bare infinitive/base verb such as List, Get, Search, Create, Update, or Retrieve; avoid noun phrases, articles, gerunds, and third-person verbs because descriptions are part of the BM25 tool-search corpus (see BM25-Friendly Descriptions below)
• stake values map to review/approval expectations

BM25-Friendly Descriptions (verb-first-mcp-tool-descriptions rule)

Tool names and descriptions both drive BM25 retrieval. Names are the strongest signal — they must be consistent and follow the verb_noun convention (e.g., list_warehouses, get_workbook). Descriptions are the secondary signal: write each one as if answering "what user intent does this tool serve?"

Rules:

1. Start with a bare infinitive verb: List, Get, Search, Create, Update, Send, Delete
2. Include platform-specific nouns that users mention in queries: warehouse, workbook, ticket, channel
3. Include common synonyms inline when the platform uses an unusual term: worksheets (sheets/tabs)
4. Include the platform name when it adds specificity (e.g., Databricks workspace, Excel workbook), but don't lead with the full brand name or repeat it redundantly across every tool
5. For platform-specific servers, avoid adding location qualifiers (e.g., in OneDrive, in SharePoint) to every tool — BM25 treats these as content tokens, so they widen the match surface and cause your tools to surface on location-based queries (e.g., a Drive search) even when the user intended a different tool

Examples:

// BAD — noun phrase, redundant "Microsoft Excel", location noise
description: "Microsoft Excel file listing from OneDrive and SharePoint."

// BAD — gerund
description: "Listing all SQL warehouses in Databricks."

// BAD — third-person verb
description: "Lists all SQL warehouses available in Databricks."

// GOOD — bare infinitive, platform noun, no location noise
description: "List all SQL warehouses available in the Databricks workspace."

// GOOD — synonym in parentheses helps BM25 match "sheets" and "tabs"
description: "Get a list of all worksheets (sheets/tabs) in an Excel workbook."

// GOOD — verb + context + common synonyms
description: "Search Slack channels, messages, and threads by keyword or topic."

Test your descriptions: add a BM25 query case (see next section) before merging. If your expected tool doesn't score > 0 in its own server-scoped index, the description is too generic or missing the key tokens the user will type.

2. Create tools/index.ts

Create front/lib/api/actions/servers/{provider}/tools/index.ts:

import { MCPError } from "@app/lib/actions/mcp_errors";
import type { ToolHandlers } from "@app/lib/actions/mcp_internal_actions/tool_definition";
import { buildTools } from "@app/lib/actions/mcp_internal_actions/tool_definition";
import { YOUR_PROVIDER_TOOLS_METADATA } from "@app/lib/api/actions/servers/your_provider/metadata";
import { Err, Ok } from "@app/types/shared/result";

const handlers: ToolHandlers<typeof YOUR_PROVIDER_TOOLS_METADATA> = {
  list_items: async ({ pageToken, maxResults }, { authInfo }) => {
    const token = authInfo?.token;
    if (!token) {
      return new Err(new MCPError("No access token provided"));
    }

    try {
      const items = [];

      return new Ok([
        { type: "text" as const, text: `Found ${items.length} items` },
        { type: "text" as const, text: JSON.stringify({ items }, null, 2) },
      ]);
    } catch (e) {
      return new Err(new MCPError("Failed to list items"));
    }
  },

  get_item: async ({ itemId }, { authInfo }) => {
    const token = authInfo?.token;
    if (!token) {
      return new Err(new MCPError("No access token provided"));
    }

    try {
      const item = {};

      return new Ok([
        { type: "text" as const, text: `Retrieved item ${itemId}` },
        { type: "text" as const, text: JSON.stringify(item, null, 2) },
      ]);
    } catch (e) {
      return new Err(new MCPError("Failed to get item"));
    }
  },

  create_item: async ({ name, description }, { authInfo }) => {
    const token = authInfo?.token;
    if (!token) {
      return new Err(new MCPError("No access token provided"));
    }

    try {
      const item = {};

      return new Ok([
        { type: "text" as const, text: `Created item "${name}"` },
        { type: "text" as const, text: JSON.stringify(item, null, 2) },
      ]);
    } catch (e) {
      return new Err(new MCPError("Failed to create item"));
    }
  },
};

export const TOOLS = buildTools(YOUR_PROVIDER_TOOLS_METADATA, handlers);

Key points:

• ToolHandlers<typeof YOUR_PROVIDER_TOOLS_METADATA> enforces one handler per metadata name and infers each handler's parameters from that tool's Zod schema
• buildTools combines metadata and handlers into ToolDefinition[]
• Keep the metadata tuple intact. Annotating it as ToolMeta[] widens names and schemas and loses per-handler parameter inference
• Generic helpers around buildTools must preserve tool names as their own generic instead of widening them to string; prefer calling buildTools directly unless a helper adds real value
• Keep the handler map local unless production code genuinely reuses it; do not export it only to make tests bypass the built tool or server surface
• access the OAuth token via extra.authInfo?.token

3. Create index.ts

Create front/lib/api/actions/servers/{provider}/index.ts:

import type { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";

import { makeInternalMCPServer } from "@app/lib/actions/mcp_internal_actions/utils";
import { registerTool } from "@app/lib/actions/mcp_internal_actions/wrappers";
import type { ToolContext } from "@app/lib/actions/types";
import { YOUR_PROVIDER_SERVER_NAME } from "@app/lib/api/actions/servers/your_provider/metadata";
import { TOOLS } from "@app/lib/api/actions/servers/your_provider/tools";
import type { Authenticator } from "@app/lib/auth";

function createServer(
  auth: Authenticator,
  toolContext?: ToolContext
): McpServer {
  const server = makeInternalMCPServer(YOUR_PROVIDER_SERVER_NAME);

  for (const tool of TOOLS) {
    registerTool(auth, toolContext, server, tool, {
      monitoringName: YOUR_PROVIDER_SERVER_NAME,
    });
  }

  return server;
}

export default createServer;

4. Register in constants.ts

Edit front/lib/actions/mcp_internal_actions/constants.ts:

• import YOUR_PROVIDER_SERVER
• add the server name to AVAILABLE_INTERNAL_MCP_SERVER_NAMES
• add the config entry to INTERNAL_MCP_SERVERS

Example:

your_provider: {
  id: 99,
  availability: "manual",
  allowMultipleInstances: true,
  isRestricted: undefined,
  isPreview: false,
  tools_arguments_requiring_approval: undefined,
  tools_retry_policies: undefined,
  timeoutMs: undefined,
  metadata: YOUR_PROVIDER_SERVER,
},

Important properties:

• id: unique stable ID, never change after deployment
• availability: manual, auto, or auto_hidden_builder
• allowMultipleInstances: true for OAuth-based integrations
• isRestricted: feature-flag or plan gating function, if needed
• isPreview: true for beta or preview integrations
• INTERNAL_MCP_SERVERS is wrapped by ensureUniqueToolNames; preserving the metadata tuple lets it reject duplicate tool names within a server at compile time

5. Register in servers/index.ts

Edit front/lib/actions/mcp_internal_actions/servers/index.ts:

case "your_provider":
  return yourProviderServer(auth, toolContext);

Optional: client.ts and helpers.ts

Use extra files when the integration grows beyond a few simple calls.

client.ts

Create a client when you need multiple API endpoints, response validation, auth header management, or retry logic.

helpers.ts

Create helpers for:

• withAuth wrappers
• response rendering
• shared data transformations

Example withAuth pattern:

import { MCPError } from "@app/lib/actions/mcp_errors";
import type {
  ToolHandlerExtra,
  ToolHandlerResult,
} from "@app/lib/actions/mcp_internal_actions/tool_definition";
import { Err } from "@app/types/shared/result";

export async function withAuth<T>(
  { authInfo }: ToolHandlerExtra,
  action: (token: string) => Promise<ToolHandlerResult>
): Promise<ToolHandlerResult> {
  const token = authInfo?.token;
  if (!token) {
    return new Err(new MCPError("No access token provided"));
  }

  try {
    return await action(token);
  } catch (e) {
    return new Err(new MCPError("Operation failed"));
  }
}

Use client.ts / helpers.ts based on complexity:

• no external API: keep everything in tools/index.ts
• 1-2 simple API calls: inline, maybe add helpers.ts
• several API endpoints: create client.ts
• complex response formatting: add dedicated rendering helpers

Alternative: function-based tools

If handlers need access to Authenticator directly, create tools through a function instead of a constant.

See front/lib/api/actions/servers/github/tools/index.ts for a full example.

Icon

Use an existing similar icon temporarily, then request the final icon from design/Sparkle and update the icon field once available.

When wiring up the icon:

1. Add the SVG to sparkle/src/logo/platforms/ if it doesn't already exist, and re-export it from sparkle/src/logo/platforms/index.ts.
2. Always check sparkle/src/logo/platforms/registry.ts and add the logo to PLATFORM_LOGOS. The marketing site resolves icons by string name via getPlatformLogo(), so a missing registry entry silently falls back to a placeholder puzzle-piece on /integrations.
3. Marketing must be redeployed for the new icon to appear on the public integrations page — explicitly tell the engineer. The change lands in the sparkle bundle that marketing builds against, so a marketing rebuild is required.

Feature Flags and Restrictions

Gate preview or limited-access servers through isRestricted in the server config, using feature flags or plan checks as needed.

Best Practices

1. Render responses for token efficiency

Always convert API responses into focused, markdown-formatted output. Avoid returning raw JSON.stringify(apiResponse) with everything the upstream API sent.

Do:

• keep only the fields the model needs
• start with a short summary
• format structured results consistently

Do not:

• return full raw API responses
• include pagination metadata or rate-limit details unless needed
• duplicate the same data in multiple formats

2. Translate errors into actionable messages

Wrap failures in meaningful MCPErrors rather than exposing raw upstream errors.

3. Choose tool stakes carefully

• never_ask: read-only operations
• low: low-impact writes
• medium: important writes
• high: destructive or high-impact actions

4. Add .describe() to schema fields

Schema descriptions help the model supply the right parameters.

5. Validate external API responses with Zod

Validate every external response to catch API drift and unexpected payloads early.

BM25 Tests

Every new server must be added to the BM25 test corpus. This is the only automated check that description quality is sufficient for tool-search retrieval.

1. Register the server in bm25_tool_search_utils.test.ts

Add an import and a SERVER_SOURCES entry. SERVERS is derived from this array and performs the Zod-to-JSON-Schema conversion used by the BM25 test corpus:

// At the top with the other imports:
import { YOUR_PROVIDER_SERVER } from "@app/lib/api/actions/servers/your_provider/metadata";

// In the SERVER_SOURCES array:
{ name: "your_provider", tools: YOUR_PROVIDER_SERVER.tools },

2. Add query cases in bm25_tool_search.test.ts

Add entries to the QUERIES array. Each entry needs:

• query: the natural-language phrase a user would type
• expected: "<server_name>.<tool_name>" — the tool that must score > 0

{ query: "list databricks warehouses", expected: "databricks.list_warehouses" },
{ query: "what sql warehouses do I have", expected: "databricks.list_warehouses" },

Tips for writing good query cases:

• Use the vocabulary users actually type, not API jargon
• Cover at least one case per tool (two is better)
• If a query is ambiguous (multiple servers could match), add maxRank: N to relax the full-corpus ranking assertion
• The single-server test ("${query}" → ${expected} is scored in ${serverName}-only index) only checks score > 0, so focus on making sure the key tokens appear somewhere in the tool's description or inputSchema

3. Run the tests

cd front
NODE_ENV=test npm test -- --reporter verbose ./lib/api/actions/servers/bm25_tool_search.test.ts

If a case fails with "Expected tool to have a non-zero score but it was not found", the query tokens don't overlap with the tool's corpus tokens. Fix the description to include the missing token, or rephrase the query to use a term that's actually in the description.

Validation Checklist

Before marking implementation complete:

• every tool defines stake, toolCostCategory, freeUsage, and displayLabels
• tool descriptions start with a bare infinitive/base verb
• tools/index.ts exists and uses ToolHandlers<typeof METADATA>
• buildTools(METADATA, handlers) builds the runtime tool definitions
• index.ts default-exports the server factory
• the server is in AVAILABLE_INTERNAL_MCP_SERVER_NAMES
• the server config is in INTERNAL_MCP_SERVERS
• the server is registered in servers/index.ts
• response rendering is implemented
• a temporary icon is set
• the icon is present in PLATFORM_LOGOS in sparkle/src/logo/platforms/registry.ts
• the engineer has been told marketing needs a redeploy for the icon to appear publicly
• feature gating is configured if needed
• server added to SERVER_SOURCES in bm25_tool_search_utils.test.ts
• at least one query case per tool added to bm25_tool_search.test.ts
• cd front && NODE_ENV=test npm test -- --reporter verbose ./lib/api/actions/servers/bm25_tool_search.test.ts passes
• cd front && npx tsgo --noEmit passes
• npm run format:changed passes from the repo root
• manual testing is complete

Troubleshooting

Server not appearing in the builder

• check availability
• check isRestricted
• verify the server name is in AVAILABLE_INTERNAL_MCP_SERVER_NAMES

OAuth connection failing

• verify the provider exists in core and front
• check client ID / secret env vars
• verify redirect URIs and scopes

Tools not working

• verify the tool is registered
• verify every published tool includes its stake
• test the API helper functions directly
• confirm authInfo.token is propagated

Type errors

• ensure the server name was added to AVAILABLE_INTERNAL_MCP_SERVER_NAMES
• run cd front && npx tsgo --noEmit

Reference Implementations

• front/lib/api/actions/servers/github/
• front/lib/api/actions/servers/snowflake/
• front/lib/api/actions/servers/google_calendar/
• front/lib/api/actions/servers/agent_sidekick_context/
• front/lib/api/actions/servers/agent_sidekick_agent_state/

Additional Resources

• MCP SDK documentation: https://modelcontextprotocol.io/
• existing server implementations in front/lib/api/actions/servers/
• legacy implementations in front/lib/actions/mcp_internal_actions/servers/
• OAuth providers in core/src/oauth/providers/