[ricardasbitinas-jpg]

• I have read CONTRIBUTING.md.
• I have created a discussion that was approved by a maintainer (for new features).

What

Adds Hostinger Router as an openai-compat provider.

• Endpoint: https://router.hostinger.com/v1
• Key: $HOSTINGER_ROUTER_API_KEY
• Defaults: claude-sonnet-5 (large), gpt-5-mini (small)
• 119 models, including the EU and US regional variants (e.g. claude-sonnet-5-eu) for data residency

Follows the Atlas Cloud pattern (#463):

• cmd/hostinger-router/main.go: generator that reads the public, unauthenticated GET /v1/models
• internal/providers/configs/hostinger-router.json: generated output
• providers.go, KnownProviders(), Taskfile.yaml (gen:hostinger-router) and update.yml, so the 6-hourly job keeps the list current

Model selection

A model is included if the catalog lists it on /v1/chat/completions, with text in/out, tools support and a context window of at least 20k. That leaves out embeddings, image, audio and Responses-only models. These are the same filters as the OpenRouter and Atlas Cloud generators.

Data source

All values come from /v1/models:

• Prices are per token in credits (1 credit = $0.01), converted to $/1M tokens. input_cache_read maps to cost_per_1m_out_cached and input_cache_write to cost_per_1m_in_cached.
• context_window comes from context_length. default_max_tokens uses the Atlas Cloud rule: half of max_output_tokens, capped at 15% of the context window.
• supports_attachments is set when architecture.input_modalities includes image.
• can_reason and low/medium/high levels are set for models whose supported_parameters include reasoning_effort (66 of 119). Claude and Phi models are left out: with reasoning_effort on chat completions, Claude replies with Anthropic finish reasons (tool_use instead of tool_calls), which stops Crush's tool loop, and the Phi reasoning models reject the parameter.

Testing

• go build ./..., go vet ./..., gofmt -l ., gofumpt -l: clean
• go test ./...: pass (including TestValidDefaultModels and TestKnownProvidersMatchesRegistry)
• golangci-lint run ./... (v2.13.1, Go 1.26.10): 0 issues
• go run ./cmd/hostinger-router/main.go: Generated hostinger-router.json with 119 models
• Crush (main, built locally) with CATWALK_URL=http://localhost:8080 against this branch's server (the task run + task run:catwalk setup): crush models hostinger lists all 119 hostinger-router/* models

Live test (2026-10-23, base URL and model IDs from the generated JSON):

# curl https://router.hostinger.com/v1/chat/completions
chat claude-sonnet-5  HTTP 200  finish=stop        content="Hi there, friend!"
chat gpt-5-mini       HTTP 200  finish=stop        content="Hi there, friend."
tool claude-sonnet-5  HTTP 200  finish=tool_calls  get_weather({"city":"Vilnius"})
tool gpt-5-mini       HTTP 200  finish=tool_calls  get_weather({"city":"Vilnius"})

$ crush run -m hostinger-router/claude-sonnet-5 "say hi"
Hi
$ crush run -m hostinger-router/gpt-5-mini "say hi"
Hi

# in an empty directory; the agent called the ls tool, then answered
$ crush run -m hostinger-router/claude-sonnet-5 "Use your ls tool to list the files in the current directory, then tell me how many there are. Do not create any files."
0 files.
$ crush run -m hostinger-router/gpt-5-mini "Use your ls tool to list the files in the current directory, then tell me how many there are. Do not create any files."
0

reasoning_effort: medium was also sent to each base model that gets reasoning levels here, plus a few regional variants. All returned HTTP 200, streaming and non-streaming.

Disclosure: I work at Hostinger. I'm happy to keep this provider's model list up to date.

🤖 Generated with Claude Code