MCP server for Vapi AI, providing tools to manage assistants, calls, phone numbers, and custom tools
The Vapi MCP server defines 13 tools with visible schemas and descriptions, but falls short of production standards across multiple dimensions. Tool naming follows verb_noun conventions well (list_assistants, create_assistant, etc.), which is a strength. However, parameter descriptions are consistently vague and incomplete. For example, create_assistant accepts an 'llm' parameter described as 'LLM configuration', lacking specifics about required fields, provider constraints, or valid model names (which ARE defined in schemas/index.ts but not surfaced in parameter descriptions). Similarly, 'toolIds' is described as 'IDs of tools to use' with no guidance on what happens if an invalid ID is passed. The voice and transcriber parameters are equally opaque. Output schemas are never documented, callers cannot know what fields to expect from create_assistant, update_assistant, or create_call responses. Error handling is absent from visible code: no guidance on what happens if an assistantId is invalid, if a rate limit is hit, or if the Vapi API returns a 4xx/5xx error. The server hardcodes limit=10 in list_tools (src/tools/tool.ts) but this is never communicated to users in the tool description, violating the enforce-result-limits pattern. Pagination is missing entirely, no offset/limit parameters on list_* calls despite potential for large result sets. Security: the VAPI_TOKEN is correctly injected server-side (src/index.ts), avoiding secrets in parameters. However, there is no audit logging visible, permission gating, or scope declarations. The Zod schemas in src/schemas/index.ts define rich validation logic (ModelProvider enums, voice providers, LLM variants) but these are not reflected in parameter descriptions, forcing LLMs to guess at valid values. For example, create_assistant's llm parameter should enumerate allowed providers (openai|anthropic|google) in its description, matching the schema. Response transformation (transformToolOutput, transformToolInput) is referenced but implementations not visible, making it impossible to verify output quality. Overall, the server is a thin wrapper around the Vapi API with basic scaffolding but insufficient documentation and error guidance for safe, reliable agent use.
Creates a new Vapi assistant
Creates a outbound call
Creates a new Vapi tool
Gets a Vapi assistant by ID
Gets details of a specific call
Gets details of a specific phone number
Gets details of a specific tool
Output schemas completely undocumented. No tool description states what fields the response contains, forcing LLMs to operate blind. Example: create_assistant returns an object but response structure (fields, types, nested objects) is never documented.
Parameter descriptions are generic placeholders. 'llm': 'LLM configuration' does not specify required fields, allowed providers (openai|anthropic|google), or valid models. Valid options ARE in Zod schemas (schemas/index.ts) but not surfaced to LLMs.
No error handling guidance. Tools do not document what happens on invalid input (bad assistantId, invalid LLM provider), rate limits, or Vapi API failures. LLMs receive no recovery instructions.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 51 | 2025-06-18+ | v2 |
| 2026-03-09 | F | 38 | - | v1 |
Lists all Vapi assistants
Lists all Vapi calls
Lists all Vapi phone numbers
Lists all Vapi tools
Updates an existing Vapi assistant
Updates an existing Vapi tool
Pagination missing. list_* tools hardcode limit=10 (src/tools/tool.ts: 'limit: 10') but do not expose offset/limit parameters to callers. If >10 items exist, LLMs cannot retrieve them. Violates enforce-result-limits pattern.
Parameter documentation incomplete for complex objects. 'voice': 'Voice configuration', 'transcriber': 'Transcription configuration', 'assistantOverrides': 'Overrides for the assistant configuration' are all vague. Zod schemas define valid providers (11labs, deepgram, azure, etc.) but LLMs cannot see them.
Idempotency not documented. Tools like create_assistant, create_call, create_tool do not state whether repeated calls with identical inputs produce idempotent results or create duplicates. Agents retry on ambiguous failures, non-idempotent tools risk duplicate records.
Transformers (transformToolOutput, transformToolInput) are referenced but implementations are not visible in source. Cannot verify that response fields are stripped of noise, that chaining IDs are included, or that output matches LLM expectations.
No permission gating or scope declarations. Tools do not declare required permissions (read:assistants, write:calls, etc.), making least-privilege configuration impossible. No audit trail of who called what.
Destructive operations (update_assistant, update_tool, create_call) lack confirmation or dry-run modes. An agent mistake could overwrite assistant instructions, tool definitions, or initiate unwanted calls with no undo mechanism.