Pi

A custom provider block in ~/.pi/agent/models.json — baseUrl, api and the model ids — puts Earendil's Pi terminal coding agent on Claude, GPT and Gemini through one key.

{
  "providers": {
    "kunavo": {
      "baseUrl": "https://api.kunavo.com/v1",
      "api": "openai-completions",
      "apiKey": "$KUNAVO_API_KEY",
      "models": [
        {
          "id": "claude-sonnet-5",
          "name": "Claude Sonnet 5",
          "reasoning": true,
          "input": ["text", "image"],
          "contextWindow": 1000000,
          "maxTokens": 128000
        },
        {
          "id": "claude-haiku-4-5",
          "name": "Claude Haiku 4.5",
          "input": ["text", "image"],
          "contextWindow": 200000,
          "maxTokens": 64000
        }
      ]
    }
  }
}

Step by step

1. Create a key at /app/keys and copy it — it is shown once.
2. Put it in the environment as KUNAVO_API_KEY. Pi resolves "$ENV_VAR" or "${ENV_VAR}" in the apiKey field, and its own warning is worth heeding: $FOO_BAR is the variable FOO_BAR, so use the braced form when literal text follows the name. A missing variable leaves the value unresolved rather than erroring loudly.
3. Create or edit ~/.pi/agent/models.json and paste the block above. A non-built-in provider needs baseUrl and an api value at either provider or model level; everything else in Pi's table is optional.
4. Start pi, run /model and pick one of the ids you declared. If they are not listed, check the key before the JSON — see the resolution-order note above.
5. Give it a task that reads and edits a real file. Pi leans on tool calling for nearly everything it does, so a first run that touches the filesystem tells you much more than a greeting does — and it is the run that would surface a streaming or tool-schema mismatch, which is exactly the class of thing Kunavo has not tested for you.

Checked against Pi's Custom Models documentation on September 21, 2026. Third-party settings move; if a field name here no longer matches what you see, that page is the authority, not this one.

Verify before you debug the client

One request settles whether a failure is the endpoint, the key, or the configuration file. If this returns JSON, the same base URL and key work in Pi.

# Settles whether a failure is the endpoint, the key, or the client.
curl -sS https://api.kunavo.com/v1/models \
  -H "Authorization: Bearer sk-kn-..."

Which model id to put in the field

Every text model is reachable as a model id — the live list is GET /v1/models, and the catalog with prices is on the models page. Rates are USD per 1M tokens, input / output.

The other route: anthropic-messages

Pi's custom-provider schema accepts four values for api — openai-completions, openai-responses, anthropic-messages and google-generative-ai — and Kunavo answers the Anthropic Messages surface as well as the OpenAI-compatible one. So the route is expressible. What this page will not do is write a base URL for it into a block you would paste, because Pi's own page shows that field two different ways and settles neither: "baseUrl": "https://proxy.example.com/v1" in its Custom Headers example, and a bare "baseUrl": "https://proxy.example.com" in its Anthropic Messages Compatibility example. Two readings, no sentence deciding between them. If you take this route, try one, and if the first call comes back 404 rather than 401, that line is the one to change.

Three more fields from the same section are worth knowing before you get there, because they exist precisely for endpoints that are not Anthropic itself:

1. compat.supportsEagerToolInputStreaming — Pi sends per-tool eager_input_streaming by default and documents setting this to false when a backend rejects the field, which makes it fall back to the legacy fine-grained tool-streaming beta header.
2. compat.supportsStrictTools — off by default for custom models; Pi says built-in Anthropic models enable it in their own metadata, so a custom Anthropic-compatible model has to set it explicitly if the endpoint accepts strict JSON-schema tool definitions.
3. compat.supportsMidConvoEffort — Pi's own guidance is narrow and worth quoting: set it “only for the exact supported Claude model on a faithful Anthropic Messages transport”. Whether this endpoint qualifies is a runtime question, and Kunavo has not answered it by running anything.

The openai-completions block at the top of this page avoids all three, which is the honest reason to start there rather than a claim that it performs better.