From the first request
to a working integration.
Start with the tested model, then choose an SDK or agent harness that supports OpenAI Chat Completions. The examples below include real requests, streaming responses and token usage.
Make your first end-to-end request.
The local prototype exposes Chat Completions under localhost:3000/v1. Your OpenAI client’s base URL should include /v1; the route adds /chat/completions automatically.
- 1
Set up your local access
Run
npm run devin this repository. The direct route requires a key whose SHA-256 hash is set inBLASCHKE_API_KEY_HASH. Use a separate, strong secret for a real deployment; the example below is for local testing only.Local development environment / shell# Generate a local hash for the development key “local-prototype”: node -e 'console.log(require("node:crypto").createHash("sha256").update("local-prototype").digest("hex"))' # Put the resulting hash in .env.local: BLASCHKE_API_KEY_HASH=<the-64-character-hash> ENABLE_LOCAL_PLAYGROUND=true # Your client then uses the corresponding plain-text key: export BLASCHKE_API_KEY="local-prototype"To call a worker, also configure
INFERENCE_BASE_URLandINFERENCE_API_KEYserver-side. Those are upstream credentials, not credentials you enter in the site. See the repository README. Hosted tenant-key management is planned, not claimed here. - 2
Choose a language
first-request.sh / shellcurl http://localhost:3000/v1/chat/completions \ -H "Authorization: Bearer $BLASCHKE_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "qwen3-8b-abliterated", "messages": [ {"role": "system", "content": "You are a concise, helpful assistant."}, {"role": "user", "content": "Explain what a Blaschke product is."} ], "stream": true, "stream_options": {"include_usage": true} }'With
stream_options.include_usage, the server also requests a terminal usage object from the upstream worker. The route preserves it for your client. In this schema,streamdefaults tofalse; set it to true as shown for streaming. - 3
Use the tested public alias
Use
minicpm-1b-uncensored. The six other catalogue aliases are planned, not enabled serving IDs. Do not select a private source filename or a Hugging Face repository path as the API model.
Match the client to the protocol.
A base URL does not convert one protocol into another. Your client and endpoint must send the same message format and, where applicable, tool-call schema.
Your agent or application
│
├─ OpenAI Chat Completions ──> Blaschke /v1/*
│ (available today)
├─ OpenAI Responses ─────────> planned adapter
│
└─ Anthropic Messages ───────> planned adapter
A changed base URL alone does not translate protocols.Current compatibility at a glance.
This matrix reflects the configured code and working worker integration, rather than capabilities inferred from a source model card.
| Capability | Current state | Details |
|---|---|---|
| Text streaming | Tested | Real OpenAI-compatible SSE token deltas. |
| Input and output usage | Tested | Counts and total, read from the inference stream. |
| Non-streaming response | Route implemented | Default stream:false; the response body and content type come from the worker. |
| Pi custom provider | Protocol compatible | Use openai-completions. The full tool-driven agent loop is not available through this text endpoint. |
| Claude Code / Anthropic Messages | Planned adapter | A different protocol is required—not just a change of base URL. |
| Codex / OpenAI Responses | Planned adapter | Responses tool schema and structured streaming need implementation and testing. |
| Large-model tool calls or images | Not verified | Source cards report some features; the local route rejects image and tool request fields. |
| Hosted accounts, billing, persistence | Not implemented | No public production account or prompt persistence is claimed. |
A verified stream on a bounded 1B CPU worker is evidence of integration, not production scale, large-model capabilities or broad refusal benchmarks. Local console inference must be explicitly enabled with its environment flag.