Ollama Pro cloud models

This guide is for Ollama’s hosted cloud models (including an Ollama Pro subscription), not a local Ollama or llama.cpp server. It records the verified integration boundary as of 2026-09-07.

No Ollama cloud account was available for this delivery. The subscription plan’s quota, rate limits, model availability for a particular account, and the resulting Orbi run were not tested. Do not read the examples below as a promise of free or unlimited capacity.

What was verified first

The installed Pi is 0.85.1. Its catalog (pi --list-models) contains no ollama provider, and pi auth check --provider ollama --json --no-refresh returns {"status":"not_ready","provider":"ollama","reason":"provider_not_found"}. Therefore:

Pi’s built-in provider and /login flow are not the Ollama path in this version.
Do not invent an Ollama entry in Pi’s OAuth credentials or assume that pi login authorizes Ollama.
Orbi must use Pi’s custom models.json-shaped pi_providers file, with Ollama’s OpenAI-compatible cloud endpoint and an API-key reference.

This conclusion is also consistent with Ollama’s official docs: local/cloud CLI use starts with ollama signin, while direct access to ollama.com creates an API key at Ollama settings/keys. Ollama’s OpenAI compatibility documentation gives https://ollama.com/v1 as the cloud base URL. Sources: Cloud, API, and OpenAI compatibility.

Prepare Ollama access

Sign in or create an account at ollama.com. If you also use the local Ollama CLI, its documented preparation command is:
ollama signin
This is not a Pi or Orbi login command.
Create an API key at ollama.com/settings/keys. The key is account secret material; never paste it into this repository, an Issue, a log, or a commit.

Put the key in the local, gitignored environment file used by Orbi/systemd:

mkdir -p .orbi
printf '%s\n' 'OLLAMA_API_KEY=replace-with-your-key' > .orbi/env
chmod 600 .orbi/env

Check the account’s current catalog instead of copying a model name from this article. Ollama documents this authenticated endpoint:
set -a; . .orbi/env; set +a curl -fsS https://ollama.com/api/tags \ -H "Authorization: Bearer $OLLAMA_API_KEY"
The response is the account/API catalog at the time of the request. The official cloud example uses gpt-oss:120b for direct API access; it is an example, not a quota or availability guarantee. A local CLI cloud tag such as gpt-oss:120b-cloud is not automatically the same model ID for the direct API.

Minimal Orbi configuration

Because Pi 0.85.1 does not have an Ollama provider, create a small local provider file in the documented Pi models.json shape. The file may be uncommitted and can contain only the model you selected from /api/tags: .orbi/pi-providers.json:

{
  "providers": {
    "ollama-pro": {
      "baseUrl": "https://ollama.com/v1",
      "api": "openai-completions",
      "apiKey": "$OLLAMA_API_KEY",
      "models": [
        {
          "id": "gpt-oss:120b",
          "name": "Ollama cloud model (verify in /api/tags)",
          "contextWindow": 131072,
          "maxTokens": 16384
        }
      ]
    }
  }
}

Replace the model id with an exact ID returned for your account. Then select it in orbi.toml:

pi_providers = ".orbi/pi-providers.json"
pi_provider = "ollama-pro"
pi_model = "gpt-oss:120b"

The key remains only in .orbi/env; systemd loads that file through EnvironmentFile. For a manual tick, export the file first:

set -a; . .orbi/env; set +a
PYTHONPATH=src python3 -m orbi.runner --config orbi.toml

Run Orbi’s config validation before claiming the provider works. It fail-fast checks the provider, exact model, endpoint, API, and referenced environment variable. Do not add the same provider to a second JSON file or to Pi’s built-in auth store.

Quota, errors, and evidence boundary

Ollama’s public docs identify cloud models and API-key authentication, but do not establish a fixed Ollama Pro quota that Orbi can promise. This delivery has no Ollama Pro account, so quota/limit behavior, plan entitlements, account-specific model visibility, and a successful Orbi Issue are 待用户自行验证 / not tested. Record only the date, plan, selected model, and redacted command/result if a user performs that test; never record the key. An authentication or quota failure can appear as an HTTP/API error from Ollama and is surfaced by Pi/Orbi as a failed model request. Do not silently retry it or call that run successful. Check the redacted log for the provider/model and the upstream status, then either correct the key/catalog or switch providers.

Smallest fallback diff

Keep the runner and task unchanged. Change only the selected provider/model and, if needed, the secret in .orbi/env:

-pi_provider = "ollama-pro"
-pi_model = "gpt-oss:120b"
+pi_provider = "local-qwen"
+pi_model = "Qwen3.8-27B"

For z.ai, Codex, or another existing template, use its exact provider/model pair from Providers and change only the corresponding pi_provider, pi_model, and local secret reference. The fallback is not a guarantee that its quota is available.

Security checklist

Keep .orbi/env and any local .orbi/pi-providers.json out of Git; use $OLLAMA_API_KEY, never the literal key, in JSON.
Do not put a key in an Issue, PR, session transcript, shell history, or diagnostic log.
Treat model IDs and HTTP status messages as safe to share only after removing authorization headers and response data.
Ollama Pro subscription validity and quotas can change; re-check the official catalog and plan terms before a production task.

Getting Started

Core Concepts

Guides

Optional Components

Releases

Ollama pro

Ollama Pro cloud models

What was verified first

Prepare Ollama access

Minimal Orbi configuration

Quota, errors, and evidence boundary

Smallest fallback diff

Security checklist

​Ollama Pro cloud models

​What was verified first

​Prepare Ollama access

​Minimal Orbi configuration

​Quota, errors, and evidence boundary

​Smallest fallback diff

​Security checklist

Ollama Pro cloud models

What was verified first

Prepare Ollama access

Minimal Orbi configuration

Quota, errors, and evidence boundary

Smallest fallback diff

Security checklist