tokenlab-deepseek-harness-provider

DeepSeek Harness lets users mount custom model providers, but TokenLab's 134 models span three native protocols while its 80 developer APIs are scattered across a full MCP profile. Setting up routing, tool namespaces, and async polling by hand is tedious and error-prone. This bundle collapses all of that into a single drop-in profile package: install it and the model picker immediately shows three mutually exclusive TokenLab routes with zero manual mapping.

Routing is driven entirely by the native format each model declares in its public TokenLab contract: 27 on OpenAI Responses, 8 on Anthropic Messages, and 99 on Chat Completions, with no substring guessing or provider-internal heuristics. On the MCP side, the official mcp-server 0.6.17 exposes 80 tools covering model discovery and pricing, multimodal generation (image, video, music, 3D), files, embeddings, rerank, translation, batches, and other allowlisted developer operations. Async media tasks are polled through the native tokenlab_wait_task tool, which propagates the caller's AbortSignal, retries only bounded transient HTTP failures, and never mutates task state.

Built for developers constructing agents on DeepSeek Harness who want TokenLab as a unified inference and multimedia backend. It is especially useful for teams that mix all three protocols in a single workflow and frequently generate images, videos, music, or 3D assets while needing cancellation safety and bounded retries.