Crewplane Ollama

Crewplane's official Ollama plugin generates text with a local model and passes the answer to downstream workflow nodes through ordinary artifacts.

Standard Crewplane installs include this distribution in the same Python 3.13+ environment. It requires Crewplane 0.3.8+. See provider setup for package installation, server configuration, and the bundled text-generation example. Generated configs include a commented local agent and plugin registration; enable both manually or follow the onboarding walkthrough. This package exports crewplane_ollama:OLLAMA; it is loaded when explicitly selected in onboarding or registered under settings.integrations.invoker.options.providers.ollama.

CLI contract

Targets Ollama v0.40.2:

ollama run MODEL --nowordwrap --hidethinking

The workflow model overrides the agent default. A nonblank model is required. The rendered prompt is sent as UTF-8 stdin; cli_cmd must begin with the configured executable and run. Both cli_cmd and extra_args are validated. Allowed literal options are --keepalive VALUE, --format VALUE, --verbose, --insecure, and --think (or --think=true|false|high|medium|low). Value options also accept =VALUE; boolean options accept =true/=false. Competing positional models/prompts, other options, managed flags, alternate prompt transport, and first-class reasoning are rejected. model_arg is ignored. Arguments retain their literal values and execution uses the runtime-supplied working directory. Hiding thinking affects displayed output, not whether thinking is computed.

Zero exit status and nonblank stdout are required. Stderr remains diagnostics; nonzero partial output, empty output, and stderr-only output fail. Logs use plain presentation. Provider token counts remain unknown. Local-server quota handling is disabled explicitly, including configured quota patterns, and failure classification does not infer quota exhaustion from generated prose. Configured ordinary retries, wall-clock timeouts, and output-idle timeouts still apply.

Availability and limits

Validation and real-run preflight probe model presence followed by diagnostic inference under one deadline from the agent's invocation_timeout_seconds, or 60 seconds when it is null. Both use the fully resolved executable, the same server environment, and the fixed prompt Reply with OK only.. Probe captures are deleted after consumption; cancellation awaits shared process cleanup. Dry-run does not probe.

Provisioning, server cloud policy, model-removal pull races, resource consumption, CLI-only cancellation, separate concurrency limits, and unchanged-version/tag identity limits are explained in the setup guide. Linux/macOS/WSL follow Crewplane's platform scope; native Windows acceptance is pending.