[nicolas-micaud]

Summary

Adds a SOKKAN Router custom endpoint guide under the LibreChat YAML AI endpoints docs, following the NEAR AI Cloud guide (#668) and the AMD Lemonade guide (#759) for the icon registration.

SOKKAN Router (https://router.sokkan.ch) is an OpenAI-compatible inference gateway operated by NINABOT Sàrl (Geneva, Switzerland). It serves open-weight models from the EU and Switzerland, with streaming and tool calls, billed per token from a prepaid balance.

• New page content/docs/configuration/librechat_yaml/ai_endpoints/sokkan.mdx: API key, endpoints.custom entry, optional X-Data-Location header, notes.
• Sidebar entry in ai_endpoints/meta.json and a card in the "Other compatible endpoints" list of ai_endpoints/index.mdx.
• Monochrome provider icon components/icons/providers/sokkan.tsx, derived from the published mark (https://sokkan.ch/brand/sokkan-icon.svg), registered in components/icons/providers/index.ts and lib/icons.tsx.

The example uses:

• SOKKAN_API_KEY
• https://router.sokkan.ch/v1
• models.fetch: false with the five models that GET /v1/models lists with "availability": "on_demand" on 2026-11-03: qwen/qwen3.6-35b-a3b-eu, qwen/qwen3.8-27b, qwen/qwen3.5-397b-a17b, openai/gpt-oss-120b, openai/gpt-oss-20b

fetch is off on purpose, and the page says why: the public model list also contains a committed-throughput model, which returns 403 committed_throughput on a pay-per-token key, and a best-effort one. With fetch: true both would appear in the model picker.

Affiliation: this guide was written by the team that operates SOKKAN Router.

Testing

• pnpm exec prettier --check on every changed file: passes.
• pnpm exec eslint --max-warnings 0 on the icon and the two icon registries: passes.
• pnpm exec fumadocs-mdx && pnpm exec tsc --noEmit: passes.
• pnpm check:locale-meta: passes (the 13 translated meta.*.json sidebars were regenerated with pnpm sync:locale-meta).
• pnpm check:config-version: passes.
• pnpm test: 42 files / 480 tests pass.
• The icon was rendered and compared with the source mark at 128 px.
• curl -fsSL https://router.sokkan.ch/v1/models | jq -e '[.data[] | select(.availability == "on_demand") | .id] | contains(["qwen/qwen3.6-35b-a3b-eu", "qwen/qwen3.8-27b", "qwen/qwen3.5-397b-a17b", "openai/gpt-oss-120b", "openai/gpt-oss-20b"])'
• Live chat completion, streaming and a tool call (get_weather → finish_reason: "tool_calls") against https://router.sokkan.ch/v1 with each of the five model ids. LibreChat itself was not run against the endpoint.

[vercel]

Someone is attempting to deploy a commit to the LibreChat's projects Team on Vercel.

A member of the Team first needs to authorize it.