Apideck MCP

Apideck MCP

The Apideck MCP server exposes the Unified API as Model Context Protocol tools. One MCP server, 330+ tools, 200+ SaaS connectors.

If you're building an agent that needs to read/write SaaS data — invoices, candidates, employees, files, tickets, customers — and your agent runtime supports MCP (Claude Code, Claude Desktop, Cursor, Windsurf, OpenAI Agents SDK, Pydantic AI, LangChain), this is usually the right choice over a language SDK.

When to prefer MCP over the language SDK

Endpoints

Hosted:    https://mcp.apideck.dev/mcp
Stdio:     npx -y @apideck/mcp
Source:    https://github.com/apideck-libraries/mcp

Pass credentials via headers (HTTP)

x-apideck-api-key:      <your Apideck API key>
x-apideck-app-id:       <your application ID>
x-apideck-consumer-id:  <end-user / customer ID>

Pass credentials via env vars (stdio)

APIDECK_API_KEY=...  APIDECK_APP_ID=...  APIDECK_CONSUMER_ID=...  npx -y @apideck/mcp

IMPORTANT RULES

• PREFER dynamic mode (?mode=dynamic or --dynamic, the default). It exposes 4 meta-tools — list_tools, describe_tool_input, execute_tool, list_scopes — instead of 330 individual tools, dropping initial token cost from ~25-40K to ~1,300. Switch to static mode only for small fixed tool sets.
• CALL list_tools first with relevant search_terms (["accounting", "invoices"], ["hris", "employees"]). Don't enumerate 330 tools.
• PREFER workflow tools over manual chains — see "Intent-grouped workflows" below. The agent that picks apideck-pay-bill instead of stitching accounting-bills-get + accounting-bill-payments-create will succeed more often and leak less context.
• HANDLE URL elicitations: when an upstream call hits a missing/expired Vault connection, the server throws UrlElicitationRequiredError (MCP error code -32042) carrying a one-time consent URL. Surface this URL to the user, wait for them to complete OAuth, then retry the original tool call.
• USE x-apideck-service-id when a consumer has multiple connections of the same unified API — e.g. the consumer connected both Xero and QuickBooks, and the user wants the call to target Xero. Without this, Apideck routes to whichever connection is first.
• USE scopes to restrict an agent to a subset of capabilities — ?scopes=read for analysis agents, ?scopes=read,write to allow non-destructive changes, omit for full access. Scopes are enforced server-side, so a "read-only" agent literally cannot call destructive tools no matter what the prompt says.

Intent-grouped workflows

The MCP server ships 4 high-level workflow tools alongside the 330 endpoint tools. Each workflow orchestrates 2-4 underlying calls behind a single intent. Prefer these when they match the user's task — agents picking the right workflow succeed more often than agents stitching the chain manually.

Workflows return a structured result with failingStep so you know which leg of the chain broke when something fails. Partial successes (e.g. one report unsupported by the connector) come back as useful data, not opaque errors.

Mode selection

The server supports three operational modes:

• Dynamic mode (default): 4 meta-tools, ~1,300 initial tokens, agent discovers via list_tools. Use this.
• Static mode: All 330+ tools listed up front, ~25-40K tokens. Use only when you're filtering to a tiny subset (e.g. ?include=accounting-invoices-list,accounting-invoices-get) and want the schemas resolved at handshake.
• Code mode (?mode=code, experimental): Agent writes JavaScript against an apideck.* SDK in a sandbox. Beta — currently uses vm.createContext which isn't production-hardened. Skip for production agents.

Coverage

330 tools across these unified APIs:

Worked example: list invoices in dynamic mode

Agent: list_tools(search_terms: ["accounting", "invoices"])
→ Returns: [
    "accounting-invoices-list",
    "accounting-invoices-get",
    "accounting-invoices-create",
    "accounting-invoices-update",
    "accounting-invoices-delete"
  ]

Agent: describe_tool_input(tool_name: "accounting-invoices-list")
→ Returns: {
    filter: { updated_since, status, ... },
    cursor: string,
    limit: number,
    "x-apideck-service-id": string,
    ...
  }

Agent: execute_tool(
  tool_name: "accounting-invoices-list",
  input: { limit: 10, "x-apideck-service-id": "quickbooks" }
)
→ Returns: { data: [...], meta: { cursors: { next: "..." } } }

Initial cost: ~1,300 tokens for the meta-tools. The agent fetches schemas only for tools it intends to call.

Related skills

• apideck-unified-api — the unified-API model itself; install this alongside apideck-mcp if your agent isn't already grounded on Apideck's routing model.
• apideck-best-practices — auth patterns, pagination, error handling, Vault — applies to MCP usage too.
• apideck-connector-coverage — how to verify a method is supported by a specific connector before calling it.
• Per-workflow playbooks: apideck-mcp-pay-bill, apideck-mcp-receive-payment, apideck-mcp-onboard-employee, apideck-mcp-month-end-close.

Reference

• Server source + docs: github.com/apideck-libraries/mcp
• Workflow tool authoring guide: docs/authoring-workflow-tools.md
• MCP protocol spec: modelcontextprotocol.io
• Glama TDQS rubric (server scores AAA): glama.ai/blog/2026-04-03-tool-definition-quality-score-tdqs