MiniMax Provider for VS Code

Community-maintained; not affiliated with, endorsed by, or sponsored by MiniMax.

Quick start

Install the extension (Marketplace or load a .vsix).
MiniMax: Set API Key (Cmd+Shift+P on macOS / Ctrl+Shift+P on Windows/Linux) — stores the key in VS Code's SecretStorage, namespaced by region.
MiniMax: Select Region — pick Global (default) or CN.
Done. Open VS Code chat, pick a MiniMax model, and start.

Everything else below is optional.

Optional settings

Setting	Type	Default	Description
`minimax.region`	enum	`global`	`global` (api.minimax.io) or `cn` (api.minimaxi.com).
`minimax.apiStyle`	enum	`anthropic`	`anthropic` (recommended) or `openai`.
`minimax.contextVariants`	object	`{}`	Per-model override; the value must match a documented variant.
`minimax.showThinking`	boolean	`false`	Stream M3's reasoning as visible text.
`minimax.promptCaching`	boolean	`true`	Cache the prompt prefix across turns. Saves cost on long sessions and agent runs.

{
  "minimax.region": "global",
  "minimax.apiStyle": "anthropic",
  "minimax.contextVariants": { "MiniMax-M3": { "tokens": 1000000 } },
  "minimax.promptCaching": true
}

Commands

Command	Description
`MiniMax: Set API Key`	Store an API key for the active region.
`MiniMax: Clear API Key`	Remove the API key for the active region.
`MiniMax: Select Region`	Quick pick between `Global` and `CN`.

Features

All 8 documented models — MiniMax-M3 (512K / 1M), MiniMax-M2.7 + highspeed, MiniMax-M2.5 + highspeed, MiniMax-M2.1 + highspeed, MiniMax-M2.
Anthropic- and OpenAI-compatible APIs (Anthropic recommended).
Tools, reasoning, and media pass-through on both APIs. M3 image input on both; M3 video on OpenAI only (use the Files API for Anthropic). M2 family rejects image/video.
Per-region SecretStorage — a Global key is never reused for CN or vice versa.
Per-model context-window override — switch M3 from 512K to 1M via minimax.contextVariants.
Prompt caching — Anthropic-style cache_control: ephemeral on the last tool, the second-to-last user message, and the last user message. Long sessions and agent runs reuse the prefix instead of re-prefilling the conversation history on every turn.
Visible thinking — opt-in minimax.showThinking to see M3's thinking_delta inline with the response.
Cancellation and error mapping — AbortController propagation; MiniMaxApiError with stable kind discriminators; authentication headers and x-api-key values redacted before any error reaches the chat UI.

API styles

Anthropic (default) — Anthropic SDK; supports streaming, system prompts, tool calls, thinking blocks, and image input (M3).
OpenAI — OpenAI SDK with reasoning_split: true so reasoning blocks stay aligned with subsequent turns; supports streaming, tool calls, and image input (M3).

The provider always reports the documented combined input+output cap as maxInputTokens = contextWindow - maxOutputTokens so requests cannot silently exceed the API limit.

Supported models

Output caps reflect the documented recommended max_tokens values that the catalog ships in src/models.ts; the documented absolute ceilings are 524,288 (512K) for M3 and 204,800 (200K) for M2.x. Under each row's combined input+output cap the recommended cap is what leaves a usable input budget — 384K on the 512K M3 variant, 140K on every M2.x variant.

Model	Default context	Output cap	Image	Tools	Reasoning
`MiniMax-M3`	512,000 (also 1,000,000)	131,072	yes	yes	yes
`MiniMax-M2.7`	204,800	65,536	text only	yes	yes
`MiniMax-M2.7-highspeed`	204,800	65,536	text only	yes	yes
`MiniMax-M2.5`	204,800	65,536	text only	yes	yes
`MiniMax-M2.5-highspeed`	204,800	65,536	text only	yes	yes
`MiniMax-M2.1`	204,800	65,536	text only	yes	yes
`MiniMax-M2.1-highspeed`	204,800	65,536	text only	yes	yes
`MiniMax-M2`	204,800	65,536	text only	yes	yes

Privacy and security

API keys live in VS Code's SecretStorage, namespaced by region.
No telemetry is collected. No request body or response content is written to disk by the extension.
API keys and x-api-key headers are redacted before any error reaches the chat UI.
Network requests go only to the documented MiniMax endpoint for the active region.

Limitations

M2 family does not accept image or video input.
M3 inline video on the Anthropic surface is not supported — use the Files API.
Token counting uses a local ceil(length / 4) estimate. The real Anthropic count_tokens endpoint is bypassed because it can hang; the local estimate is good enough for VS Code's prompt-fit check, and an oversized prompt is rejected by the upstream API with a clear error anyway.
No live API key in CI; the test suite uses stub SDK clients. Run a live smoke test before shipping.
Reasoning content is preserved across turns but is not surfaced to the chat UI by default.

Logo attribution

The extension icon (resources/icon.png) is derived from the MiniMax logo. MiniMax and the MiniMax logo are trademarks of MiniMax. The icon is used only to indicate MiniMax model-provider compatibility; no endorsement by MiniMax is implied.

Licensing & third-party notices

AGPL-3.0. See LICENSE. Bundled third-party dependency licenses and attributions are in THIRD-PARTY-NOTICES.md.

MiniMax Provider

devparanjay