Video Providers

Video Context MCP supports six AI providers for video analysis. You can configure all of them and the server will fall back automatically when one is unavailable or rate-limited.

Provider Comparison

Feature	Gemini 3.5 Flash	MiniMax-M3	Kimi K2.6	Qwen3.7	MiMo-V2.5	GLM-4.6V
Price	Free tier	Paid	Paid	Paid	Paid	Free tier
Context window	1M tokens	1M tokens	256K tokens	1M tokens	256K tokens	128K tokens
Native video upload	Up to 2 GB	No — image-only	Up to 100 MB	No	No	No
Frame/image cap	n/a	100 images per request	n/a	20 images per request	20 images per request	20 images per request
Large local files	Native upload	Always frame-based	Native upload up to 100 MB	S3 relay or frame fallback	S3 relay or frame fallback	S3 relay or frame fallback
Best for	Default free provider	Long-video understanding	Broadest native format support	1M-context paid fallback	Thinking-mode workflows	Free last-resort fallback

Gemini 3.5 Flash — Default

The default provider. Free with no credit card required. Accepts files up to 2 GB natively. An excellent all-round provider with the largest context window.

Environment variable: GEMINI_API_KEY
Get API key →

MiniMax-M3

MiniMax-M3 is a paid, image-only provider built for long-video understanding. The server extracts keyframes and sends them as images, with support for up to 100 images per request and 10 MB per image. It is the second provider in the fallback chain and the strongest frame-based option for long videos.

Environment variable: MINIMAX_API_KEY
Optional timeout override: M3_REQUEST_TIMEOUT_MS
Get API key →

Kimi K2.6

Paid provider with the broadest native video format support. Accepts files up to 100 MB natively without S3 relay, which is why it now sits ahead of Qwen in the fallback chain.

Environment variable: MOONSHOT_API_KEY
Get API key →

Qwen3.7

Paid provider with a 1M-token context window. Local files use base64 with a practical cap of about 10 MB unless you configure the S3 Relay.

Environment variable: DASHSCOPE_API_KEY
Optional endpoint override: QWEN_BASE_URL
Get API key →

MiMo-V2.5

Paid provider with Xiaomi’s thinking-mode support. Local files use base64 with a practical cap of about 10 MB unless you configure the S3 Relay.

Environment variable: MIMO_API_KEY
Get API key →

GLM-4.6V

Free tier available with no credit card required. Uses base64 video encoding for local files (cap: ~12 MB); falls back to S3 relay or frame-based analysis for larger files. In v1.3.0 it moved to the end of the fallback chain because its free tier is the most aggressively rate-limited.

Environment variable: Z_AI_API_KEY
Get API key →

Fallback Chains

When a provider fails or its key is missing, the server automatically tries the next provider in priority order:

Default Provider	Fallback Order
Gemini (default)	Gemini → MiniMax-M3 → Kimi → Qwen → MiMo → GLM
MiniMax-M3	MiniMax-M3 → Gemini → Kimi → Qwen → MiMo → GLM
Kimi	Kimi → Gemini → MiniMax-M3 → Qwen → MiMo → GLM
Qwen	Qwen → Gemini → MiniMax-M3 → Kimi → MiMo → GLM
MiMo	MiMo → Gemini → MiniMax-M3 → Kimi → Qwen → GLM
GLM	GLM → Gemini → MiniMax-M3 → Kimi → Qwen → MiMo

Change the Default Provider

Set the VIDEO_MCP_DEFAULT_PROVIDER environment variable in your MCP config:

"VIDEO_MCP_DEFAULT_PROVIDER": "m3"

Valid values: gemini (default), m3, kimi, qwen, mimo, glm

To ensure the LLM respects your configured default provider, add the following to your Copilot instructions (.github/copilot-instructions.md) or AGENTS.md:

If VIDEO_MCP_DEFAULT_PROVIDER is set in the MCP config, respect that value as the default provider for video tools. Do not specify a provider parameter unless the user explicitly requests one. :::

Large Local Files

GLM, Qwen, and MiMo all have a ~10–12 MB base64 encoding limit for local files. Kimi accepts native upload up to 100 MB. MiniMax-M3 is image-only, so it always analyzes extracted frames and never counts as an upload-capable fallback. When a GLM, Qwen, or MiMo file exceeds the base64 limit:

The server first tries to fall back to an upload-capable provider (Gemini or Kimi) if available in the chain.
If that fails too, the server falls back to frame-based analysis (evenly-spaced keyframes) as a last resort.
For the best quality with large local files, configure the S3 Relay — GLM, Qwen, and MiMo will receive a presigned URL to the full video, bypassing the limit entirely.