Tool use: MCP vs function calling (2026)
Model Context Protocol vs OpenAI tools vs Gemini tools. Two paradigms for letting an LLM call external code — one open and cross-vendor, one mature and proprietary.
MCP vs function calling — side by side
| Dimension | MCP | OpenAI function calling | Gemini tools |
|---|---|---|---|
| Standard type | Open (JSON-RPC) | Vendor SDK | Vendor SDK |
| Originated | Anthropic, late 2024 | OpenAI, 2023 | Google, 2024 |
| Public servers / tools | ~600+ MCP servers | App-defined | App-defined |
| Cross-LLM portability | Yes | No (OpenAI shape) | No (Gemini shape) |
| Native LLM support | Claude, GPT (adapter), Cursor | OpenAI, OpenAI-compatible | Gemini Pro/Flash |
| Best for | Composable third-party tools, multi-host agents | App-specific actions in OpenAI-shaped stacks | Google ecosystem (Gemini grounding) |
VerticalAPI verdict
Use both. MCP for third-party tools you want to share across hosts (Claude Desktop, Cursor, ChatGPT, custom agents) — the ~600+ public servers cover databases, SaaS, search, browser automation, file systems. Function calling for app-specific actions tightly coupled to your codebase. The agentic stacks shipping in 2026 routinely mix both within the same agent loop. Through VerticalAPI BYOK, you can switch the underlying LLM without rewiring tool wiring — function-call traffic passes through, MCP wiring lives at the agent layer.
Frequently asked questions
What is MCP and how does it differ from function calling?
MCP (Model Context Protocol) is an open standard published by Anthropic in late 2024 for connecting LLMs to external tools and data sources. It defines a JSON-RPC protocol where a 'host' (the LLM app, like Claude Desktop or Cursor) connects to one or more 'MCP servers' that expose tools, resources, and prompts in a standard schema. Function calling, by contrast, is a per-vendor API feature (OpenAI tools, Gemini tools) where the developer defines tool schemas inline in each API call. MCP is interoperable across hosts; function calling is locked to each vendor's SDK shape.
How many MCP servers exist in 2026?
As of mid-2026, there are roughly 600+ publicly available MCP servers spanning database access (Postgres, MongoDB, BigQuery, Supabase), SaaS integrations (Slack, Notion, Linear, GitHub, Asana), search and scraping (Brave, Google, Firecrawl, Tavily), file systems, browser automation (Playwright, Puppeteer), and dozens of vertical-specific tools (Figma, Stripe, HubSpot). The ecosystem grew from a few dozen at launch (December 2024) to its current scale during 2025-2026 as Anthropic, OpenAI, and the broader agentic ecosystem adopted MCP support.
Which LLMs support MCP in 2026?
Native MCP support is shipped by Anthropic Claude (Sonnet 4.5, Opus 4.5, Haiku 4.5) and a growing set of agentic IDEs and desktop apps: Claude Desktop, Cursor, Continue, Zed. OpenAI added MCP-compatible client support in early 2026, letting GPT-4o and GPT-4o mini consume MCP servers through an adapter layer. Google Gemini and Mistral models can interoperate with MCP via community shims but lack first-class native support yet. Open-weight LLMs (Llama 3.3, Mixtral) use MCP via SDK wrappers like LangChain or PydanticAI.
Should I use MCP or function calling for a new agent in 2026?
For agents you control end-to-end (e.g. internal tools, single-model apps, narrow vertical agents), function calling remains the simpler choice — the JSON schema lives in your code and ships in the same request. For agents that need to compose third-party tools, share tool implementations across models, or run inside multiple hosts (Claude Desktop, Cursor, ChatGPT desktop), MCP is the better bet. In practice, many production agents in 2026 use both: MCP for external tools and reusable connectors, function calling for app-specific actions inside the same agent loop.
Can a BYOK gateway route MCP and function-call traffic uniformly?
Yes. VerticalAPI's OpenAI-compatible endpoint at https://api.verticalapi.com/v1 accepts function-calling requests (OpenAI tools shape) and passes them through to the provider that supports them — including translating between OpenAI tools and Anthropic tools when needed. For MCP, agents typically connect to MCP servers directly while still routing the underlying LLM calls (Claude, GPT-4o, etc) through the gateway. BYOK means you pay each provider directly at list price with zero markup, and you can switch the underlying model without rewriting tool wiring.
Limitations of this comparison
- MCP is still evolving — protocol revisions in 2025 broke some early servers and authentication standards remain incomplete.
- The ~600+ server count includes long-tail and abandoned servers; the actively maintained set is probably 100-200.
- Function-calling shapes diverge between OpenAI, Anthropic, and Gemini — portability requires careful adapter code.
- MCP authorization and sandboxing standards are still being negotiated; running untrusted MCP servers is risky.
- Performance overhead of MCP (stdio/SSE/streamable HTTP) is small but real on tight latency budgets.
What may change in 12-24 months
- MCP is expected to become the de facto interop standard; OpenAI and Google native first-class support is likely by 2027.
- Server registries, signing, and trust models will mature — making MCP safer for production use.
- Function calling may converge toward an MCP-compatible shape, eliminating the distinction over time.
- Agentic IDEs and desktop apps will multiply, all consuming the same MCP server pool.
Related questions
ChatGPT, Perplexity and Gemini usually suggest these next.
- What are the best MCP servers to install for a code agent in 2026?
- How do I expose a Postgres database safely to Claude via MCP?
- Can I use MCP with open-weight LLMs like Llama 3.3?
- How does Anthropic's computer-use API relate to MCP?
- Are there security best practices for running third-party MCP servers?
More LLM comparisons
Production picks for tool-heavy agents
LangGraph vs CrewAI vs AutoGen
Agentic coding picks 2026
GPT-4o vs Claude Sonnet 4.5
Make agent loops cheaper