Providers

mew supports multiple LLM providers. Each provider has a “shape” that determines how mew communicates with it. You can use built-in providers, add custom ones, or wrap multiple providers behind a router.

Built-in providers

These are configured out of the box. Set the credential and you’re ready.

Provider	Shape	Default model	Credential env var
opencode-zen	openai	deepseek-v4-flash	MEW_CRED_OPENCODE_ZEN
opencode-go	openai	-	MEW_CRED_OPENCODE_ZEN
z-ai	openai	-	MEW_CRED_Z_AI
deepseek	openai	deepseek-v4-flash	MEW_CRED_DEEPSEEK
umans	anthropic	-	MEW_CRED_UMANS
alibaba-token-plan	openai	-	MEW_CRED_ALIBABA_TOKEN_PLAN
alibaba-token-plan-cn	openai	-	MEW_CRED_ALIBABA_TOKEN_PLAN_CN

opencode-zen and opencode-go share the same credential. umans uses the Anthropic shape and hits /v1/messages with x-api-key headers. The Alibaba Token Plan providers are OpenAI-compatible (/compatible-mode/v1); alibaba-token-plan-cn targets the mainland China endpoint and has its own credential.

Provider shapes

Two shapes are supported:

• openai: SSE, delta-based streaming. Used by OpenAI-compatible APIs (DeepSeek, OpenAI, openai-compatible gateways). Streams choices[0].delta.content as text deltas.
• anthropic: SSE, content-block events. Used by Anthropic-compatible APIs (Claude, Z.AI, Umans). Uses named events like content_block_start and content_block_delta. Thinking blocks become Part::Reasoning.

The shape determines how mew parses the SSE stream and maps it to ProviderEvent variants. See Adding a Provider for implementation details.

Adding a custom provider

Add a [providers.<id>] section to your config.toml:

[providers.my-provider]
shape = "openai"
base_url = "https://api.example.com/v1"
credential_ref = "MY_API_KEY"
model = "my-model-id"

Then set the credential:

export MEW_CRED_MY_API_KEY="sk-..."

The credential_ref is resolved through env vars, keyring, or credentials.json. See Configuration for the resolution order.

Router provider

The router wraps two providers (small + big) behind the same Provider trait with automatic switching. It picks the cheap model for simple turns and switches to the capable model when tool calls appear or the conversation exceeds a turn threshold.

[providers.smart]
shape = "openai"
kind = "router"
small = "deepseek/deepseek-v4-flash"
big = "z-ai/glm-4.5-air"
credential_ref = "deepseek"

How the router decides

The router starts with the small model. It switches to the big model when either:

• Tool calls appear in the response: the small model started generating tool calls, so the router hands off to the big model for the rest of the turn.
• Turn threshold exceeded (default 3): after 3 turns in one agentic loop, the router switches to the big model for robustness.

The Routed wrapper preserves the display model name so the TUI status line shows what you chose, even though the actual model may differ per turn. A single session can bounce between models without you knowing.

Model catalog

mew includes a built-in model catalog (from models.dev with 24-hour cache) that provides pricing, context windows, capabilities, and thinking variant defaults. The catalog is used to:

• Populate the model picker (Ctrl+P and select “Switch Model”)
• Resolve thinking variant names to provider-specific params
• Show pricing in the /cost command
• Set context windows for compaction thresholds

You can override or extend the catalog with [[models]] entries in config.toml. See Configuration for the format.

Thinking variants

Some models support configurable thinking/reasoning levels. When you set a thinking variant, mew translates it into the provider-specific parameters the model expects.

Available variants depend on the model:

Model	Variants
DeepSeek V4	high, max
GLM 5.2	high, max
GLM 5/5.1	thinking
GPT-5 family	minimal, low, medium, high
Claude Opus 4.7+ / Fable	low, medium, high, xhigh, max
Claude Sonnet 4.6	low, medium, high, max
MiMo v2.5	thinking
MiniMax M3	thinking, none
Grok-3-mini	low, high

Note that if you use a supported provider that provides thinking variants (e.g. umans), we will use those.

Set a variant with:

/thinking high

Or open Ctrl+P and select “Thinking Variant”. Press Ctrl+P repeatedly to cycle through the available options for the current model.

To disable thinking:

/thinking off

When no variant is set, mew uses the catalog default for the model.