blaschkeTry the console

Connect the model to
your existing workflow.

Use an OpenAI-compatible client directly, or keep another harness in place and add the appropriate protocol adapter. The sections below explain the practical difference between those two paths.

OpenAI-compatible clients

Supported today

Text chat via OpenAI Chat Completions. SDK examples and a Pi provider configuration are included below.

[ ]

Other protocols

Adapters planned

Claude Code and Codex do not become compatible just because an endpoint has /v1 in its URL.

Register Blaschke as a custom Pi provider.

Pi supports custom models through its provider registry. Use an openai-completions model definition, the public alias, and a context window that matches the tested worker.

Config path~/.pi/agent/models.json
~/.pi/agent/models.json / json
{
  "providers": {
    "blaschke": {
      "baseUrl": "http://localhost:3000/v1",
      "apiKey": "$BLASCHKE_API_KEY",
      "api": "openai-completions",
      "models": [
        {
          "id": "qwen3-8b-abliterated",
          "name": "Blaschke Qwen3 · 8B Abliterated v2",
          "reasoning": false,
          "input": ["text"],
          "cost": {
            "input": 0.18,
            "output": 0.18,
            "cacheRead": 0,
            "cacheWrite": 0
          },
          "contextWindow": 4096,
          "maxTokens": 512
        }
      ]
    }
  }
}
  1. 01

    Set the base URL and key reference

    baseUrl includes /v1. The $BLASCHKE_API_KEY reference resolves from Pi’s environment when it starts; it is not the literal key submitted to the worker.

    Shell / shell
    export BLASCHKE_API_KEY="local-prototype"
    # Your server must contain that key’s SHA-256 hash.
    # For a future hosted deployment:
    # export BLASCHKE_API_KEY="<your-blaschke-key>"
  2. 02

    Choose the provider and model

    Set your preferred provider and alias with /model, or use selectors supported by your installed Pi version.

    Pi CLI selector / shell
    pi --provider blaschke --model qwen3-8b-abliterated
  3. 03

    Start with text tasks

    This is a provider-registration example, not a claim that the full Pi agent loop is tested. Context and generation limits follow the selected model: MiniCPM has 4,096 tokens of context and a 512-token output cap; Qwen3-8B has a 16,384-token serving context and a 2,048-token output cap. The console sends up to four recent exchanges.

    The current request schema does not accept tool definitions, so tool-driven coding sessions need a dedicated capable model route and tested protocol before they can be recommended.

This local prototype’s Pi provider example follows the installed @earendil-works/pi-coding-agent documentation for the openai-completions, apiKey, input, cost, contextWindow and maxTokens field names. Consult the documentation for your installed Pi version before adding other options.

Use an Anthropic-compatible route.

Planned adapter

Claude Code can use alternate providers, but its harness uses Anthropic Messages. The supported Blaschke endpoint receives OpenAI Chat Completions. Setting ANTHROPIC_BASE_URL to this service today is not a working integration.

Future routing pattern (not runnable against the current server) / shell / planned, not active
# Only after an Anthropic Messages adapter is deployed and tested:
export ANTHROPIC_BASE_URL="https://your-blaschke-host.example/api/anthropic"
export ANTHROPIC_AUTH_TOKEN="your-blaschke-key"
export ANTHROPIC_MODEL="minicpm-1b-uncensored"
claude

The future adapter must translate system messages, SSE event names, stopping reasons and the requested models. An Anthropic-shaped route alone is still not proof that source-reported tools, large contexts or reasoning controls work.

Official Claude Code LLM gateway documentation

Use an OpenAI Responses-compatible route.

Planned adapter

Codex can route through custom model providers. Its harness uses OpenAI Responses, while this prototype accepts Chat Completions. Do not point an unchanged provider config at the current server and expect a working substitution.

Future ~/.codex/config.toml pattern (not an available route) / toml / planned, not active
# Conceptual pattern, after a Responses adapter is deployed:
model_provider = "blaschke"
model = "minicpm-1b-uncensored"

[model_providers.blaschke]
name = "Blaschke"
base_url = "https://your-blaschke-host.example/api/openai/v1"
env_key = "BLASCHKE_API_KEY"
wire_api = "responses"

Responses includes its own structured event stream and tool schema. A production Codex integration must retain that schema rather than stripping tools into plain text. The prototype text endpoints are not substitutes for verified tool support and harness compatibility.

Official Codex provider configuration

Configure the fields your harness expects.

For SDKs and generic harnesses, set the values below and retain your existing request-building logic. The wire protocol must match the endpoint you intend to call.

Base URLhttp://localhost:3000/v1
API keylocal-prototype in this local example; configure its hash server-side
Modelqwen3-8b-abliterated
Wire protocolopenai-completions / OpenAI Chat Completions
ContentsSystem / user / assistant messages, plain text
UsageFinal streaming chunk, when requested

If a harness expects Anthropic Messages, Responses, embeddings, arbitrary models or a tool-calling schema, a text-only route is not enough. Implement and test those exact interfaces first.