Connect Your MCP Server With MiniMax M3 & M2.7 (2026 Guide)

You can connect any MCP server to MiniMax M3 in about 60 seconds — paste the server URL into MCP Agent Studio, pick MiniMax M3 (or M2.7 / M2.5) from the model dropdown, and start chatting. Every tool call shows live, no MiniMax key required.

MiniMax M3 landed on June 1, 2026 and pushed MiniMax\u2019s open-weight stack into frontier territory for agents — strong coding, native multimodality, and a 1M-token window most closed models can\u2019t match. If your last look at connecting an MCP server with MiniMax was on the M1 reasoning model, the agentic story is far better now.

This post covers the full lineup, why these models suit MCP, three ways to wire a server up to them, MiniMax\u2019s own MCP server, and the pitfalls that cost me the most time.

The MiniMax Model Lineup for MCP (M3, M2.7, M2.5, M1)

There are two families that matter for tool calling: the M2 series (M2 through M2.7) and the new M3. The older M1 is still worth knowing about.

Model (API id)	What it is	Context	Tool calling?
MiniMax-M3	Newest flagship — open-weight, natively multimodal, computer-use, agentic	1M tokens (min 512K guaranteed)	Yes — agentic tool use, OpenAI + Anthropic SDK
MiniMax-M2.7	Highest-intelligence M2 model	Long context	Yes
MiniMax-M2.5	Tuned for tool calling and search	Long context	Yes
MiniMax-M2.1	Agent-tuned M2 variant	Long context	Yes
MiniMax-M2	Base M2 — MoE, 230B total / 10B active	Long context	Yes
MiniMax-M1	Open-weight hybrid-attention reasoning model	1M tokens (native)	Yes — structured function calling

Each M2 model also ships a -highspeed variant (for example MiniMax-M2.7-highspeed) that trades a little quality for lower latency — handy for high-volume agents.

On the Artificial Analysis Intelligence Index, MiniMax reports M2.7 at 50, M2.5 at 42, and M2.1 at 39 — so M2.7 is the strongest M2 model, and M3 sits above the series for agentic and multimodal work.

Why MiniMax Models Work Well for MCP

MCP tools reach a model as function definitions. So any model with solid function calling can drive an MCP server. MiniMax checks three boxes here.

1. OpenAI-Compatible Function Calling

MiniMax\u2019s chat endpoint takes the same tools array you already use with GPT — type: "function" plus a JSON-Schema parameters block. When the model decides to act, it returns tool_calls and a finish_reason of tool_calls.

That matters because MCP\u2019s tools/list maps one-to-one onto that tools array. No custom adapter — the same bridge that wires GPT to MCP wires MiniMax to MCP.

2. A 1M-Token Context on M3 and M1

Long agent loops accumulate tool output fast. M3\u2019s MiniMax Sparse Attention and M1\u2019s native 1M window mean a multi-step MCP agent can keep every tool result in context without truncation.

For workloads like reading a whole repo through a GitHub MCP server, or scanning a long Postgres schema, that headroom is the difference between one clean run and constant re-summarising.

3. Native Multimodality and Computer-Use on M3

M3 reads images and video and supports computer-use, not just text. So an MCP agent built on M3 can take a screenshot tool\u2019s output and reason about it directly — a step most text-only models can\u2019t take.

Connect your MCP server with MiniMax free →

Connect Your MCP Server With MiniMax — 3 Ways

Three ways to wire a server up to MiniMax, ordered by setup time. Option 1 takes about 60 seconds and needs no code.

Option 1 — MCP Playground Agent Studio (60 seconds, no code)

This is the path I\u2019d pick for most people. MCP Agent Studio handles the model-to-MCP bridge for you, ships MiniMax M3, M2.7 and M2.5 in the model dropdown, and runs the whole agent loop in the browser. No SDK, no MiniMax API key.

Step-by-step:

1. Open /mcp-agent-studio and sign in (free credits are added to your account).
2. In the MCP Servers panel, click Add server. Paste your server URL — Streamable HTTP, SSE, or HTTP. Add a bearer token in the Headers field if needed.
3. Click Connect. Agent Studio runs tools/list and shows every tool it found. If the count looks right, you\u2019re wired up.
4. In the Model dropdown, pick MiniMax M3 (hardest reasoning + multimodal), M2.7 (top text quality), or M2.5 (tool-calling + search, lower cost).
5. Send your first prompt. Every tool call, argument, and result shows inline as the agent runs.

No MCP server yet? Head to /mcp-hosted and deploy one in a click — Postgres, Stripe, GitHub, MongoDB, Playwright and 35+ more. You get a live HTTPS URL plus bearer token for step 2.

Option 2 — OpenAI SDK + MiniMax Endpoint

Building your own runtime? MiniMax\u2019s chat API is OpenAI-compatible. Change the base URL and the model name and existing function-calling code routes to MiniMax.

from openai import OpenAI

client = OpenAI(
    api_key="...",
    base_url="https://api.minimax.io/v1",
)

response = client.chat.completions.create(
    model="MiniMax-M3",            # or MiniMax-M2.7 / MiniMax-M2.5
    messages=[{"role": "user", "content": "List my open GitHub PRs"}],
    tools=mcp_tools_as_openai_functions,   # MCP tools/list -> OpenAI tools[]
    tool_choice="auto",
)

You still bridge MCP\u2019s tools/list into the tools[] array, parse each tool_calls entry into an MCP tools/call, feed the result back, and loop. Agent Studio does all of this for you.

Option 3 — Anthropic SDK (Recommended for M3)

MiniMax recommends the Anthropic SDK for M3 specifically — point the base URL at MiniMax, set the model to MiniMax-M3, and use the standard Anthropic tools format. Any MCP client built around Claude\u2019s tool-use loop then works unchanged.

This is the cleanest path if your stack already speaks the Anthropic Messages API. M3 also accepts plain HTTP requests if you don\u2019t want an SDK at all.

The Other Direction: MiniMax\u2019s Own MCP Server

\u201CConnect MiniMax with MCP\u201D has a second meaning worth clearing up. MiniMax ships an official MCP server — MiniMax-MCP — that exposes MiniMax\u2019s media APIs as MCP tools to any client.

So Claude, Cursor, Windsurf, Cherry Studio, or the OpenAI Agents SDK can call MiniMax to:

• Text-to-speech and voice cloning (plus a voice_design tool that builds a custom voice from a text prompt)
• Image generation (image-01)
• Video generation (MiniMax-Hailuo-02)
• Music generation (music-1.5)

It ships in Python (stdio, SSE) and JavaScript (stdio, REST, SSE), configured with your MiniMax API key and base path. There\u2019s also MiniMax-Coding-Plan-MCP — a coding-focused server with AI search and vision analysis tools.

Quick way to keep it straight: in this direction, MiniMax is the tool provider. In the three options above, a MiniMax chat model is the brain calling the tools.

Hands-On: 3 MCP Servers Against MiniMax M3

Three common MCP servers, each with a prompt needing at least three tool calls. I ran every prompt through Agent Studio with M3 selected.

Test 1: GitHub MCP (Search + Read)

Result: M3 called search_issues, then fanned out get_issue_comments across the matches, and grounded its summary in the actual comment text. The 1M window meant it never dropped an earlier issue from context.

Test 2: Postgres MCP (Analytical SQL)

Result: M3 grounded itself with list_schemas and describe_table before writing a window-function query. Clean tool arguments across all runs — no malformed JSON on the quoted SQL identifiers.

Test 3: Multi-MCP — GitHub + Slack Together

Result: Two MCP servers connected at once. M3 picked the right tool from each without me labelling them, then composed a tidy Slack message. M2.7 handled the same task on text quality; I reached for M3 when the prompt involved a screenshot.

MiniMax vs Claude, GPT-5, DeepSeek on MCP Workloads

Where MiniMax fits, based on the published specs and my own runs:

• M3 is the open-weight pick for multimodal agents. Its 83.5 BrowseComp edges Claude Opus 4.7\u2019s 79.3, and computer-use plus image input is rare in open weights.
• M2.7 is the value play for text-only MCP agents — strong tool calling without M3\u2019s multimodal overhead.
• M2.5 leans into tool calling and search, at lower credit cost — good for high-volume agents.
• Claude is still a touch more polished on natural-language final answers; for tool execution itself, M3 holds its own.

The honest move is to run the same prompt across MiniMax, Claude and GPT-5 side-by-side and read the tool-call traces. That\u2019s exactly what Compare mode in Agent Studio is for.

Pitfalls I Hit With MiniMax + MCP

1. Pick a Chat Model, Not a Media Model

Pointing an MCP client at image-01 or speech-2.8-hd won\u2019t call your tools — those are generation models behind MiniMax\u2019s own MCP server. Use MiniMax-M3, MiniMax-M2.7, or MiniMax-M2.5.

2. M1 Reasons, But It\u2019s Not the Agentic Flagship

M1 supports function calling and a 1M context, but it\u2019s a reasoning model from an earlier generation. For new agent builds, M3 and the M2.7/M2.5 pair are the better default.

3. Promo Pricing Is Temporary

M3 launched at promo pricing around $0.30/M input and $1.20/M output. Model your cost projections on list price, not the launch promo, before you commit a production agent.

4. Match the SDK to the Model

M3 is happiest on the Anthropic SDK (MiniMax\u2019s own recommendation); the M2 series runs fine on the OpenAI SDK. If tool calls misbehave, try the other compatible SDK before assuming the model can\u2019t do it.

How MCP Playground Helps

• 40+ models in one UI — MiniMax M3, M2.7, M2.5 plus Claude Opus 4.7 / Sonnet 4.6, GPT-5.4, Gemini 3.1 Pro, DeepSeek V4, Grok, Qwen. No MiniMax API key needed.
• Paste any MCP URL — works with hosted GitHub, Linear, Vercel, Supabase, Datadog, or your own remote.
• One-click hosted MCP servers — 40+ pre-configured: Postgres, Stripe, MongoDB, Playwright, Brave Search and more. Deploy in under a minute.
• Compare mode — run one prompt across M3, Claude and GPT-5 side-by-side, with full tool-call traces.

Bottom Line

MiniMax M3 is a genuine open-weight option for MCP agents — 1M context, native multimodality and computer-use, and an OpenAI/Anthropic-compatible API that drops into any existing MCP client. For text-only agents, M2.7 and M2.5 cover the value end.

Keep the two directions straight: MiniMax chat models drive your MCP servers, and MiniMax\u2019s own MCP server exposes its media APIs to other clients.

The fastest way to validate it on your MCP server is Agent Studio — pick MiniMax M3, M2.7 or M2.5, paste the URL, and run your hardest prompt three times.

Related Reading

• What Is Model Context Protocol? A Plain-English Guide
• Best AI Model for MCP Tool Calling
• Connect Your MCP Server With DeepSeek V4
• Test Your MCP Server With Qwen Models
• Hosted MCP Server: 7 Real Use Cases

Frequently Asked Questions

Which MiniMax models support MCP / tool calling?

The chat models all support function calling, which is how MCP tools get invoked: MiniMax-M3, MiniMax-M2.7, MiniMax-M2.5, MiniMax-M2.1, MiniMax-M2, and the open-weight reasoning model MiniMax-M1. MiniMax\u2019s media models (image, speech, video) do not call tools.

When was MiniMax M3 released and what can it do?

June 1, 2026. M3 is an open-weight, natively multimodal agentic model with a 1M-token context (MiniMax Sparse Attention) and computer-use support. Reported benchmarks: 59% SWE-Bench Pro, 66% Terminal-Bench 2.1, and 83.5 BrowseComp.

How do I connect a MiniMax model to my MCP server?

Easiest path: paste the MCP server URL into MCP Agent Studio and pick MiniMax M3, M2.7, or M2.5 from the model dropdown — no key needed. For code, MiniMax\u2019s chat API is OpenAI-compatible at https://api.minimax.io/v1/chat/completions (M3 also supports the Anthropic SDK). Map MCP tools/list into the tools[] array and loop on tool_calls.

Is MiniMax\u2019s MCP server the same as using MiniMax models with MCP?

No — they\u2019re opposite directions. MiniMax\u2019s own MCP server (MiniMax-MCP) exposes its TTS, voice cloning, image, video and music APIs as tools for any MCP client. Using a MiniMax chat model with MCP means the model is the brain that calls tools from your servers.

M3 vs M2.7 vs M2.5 — which should I use for MCP agents?

Use M3 for the hardest reasoning, multimodal input, or computer-use. Use M2.7 for top text quality on text-only agents. Use M2.5 for tool-calling and search at lower cost on high-volume workloads. Each M2 model also has a -highspeed variant for lower latency.

How do I test MiniMax against an MCP server without writing code?

Use MCP Agent Studio. Paste the MCP server URL, pick MiniMax M3, M2.7 or M2.5 from the dropdown, and start chatting. Every tool call and result shows live. Free credits on sign-up — no MiniMax API key required.