Use a local Ollama model from Claude Code to review code or delegate tasks.

This plugin lets you run code reviews and background rescue tasks against an Ollama server — local by default, with optional access to hosted frontier models via Ollama Cloud (:cloud suffix). No OpenAI account, no API key plumbing, and your code stays on your hardware unless you explicitly choose a cloud-hosted model.

1. Install Ollama — download the desktop app or follow the CLI instructions at ollama.com.

2. Pull a model:

   ollama pull llama3.1:8b

3. Install the plugin (placeholder — update once published to a marketplace):

   /plugin install ollama@darrylmorley/ollama-plugin-cc

4. Run setup:

   /ollama:setup

   Setup checks that Ollama is installed, running, and has at least one model. It also lets you set a default model and optionally enable the stop-time review gate.

5. Try a review:

   /ollama:review

See the ollama-model-prompting skill for full guidance. Short version:

Empirically battle-tested against a SQL-injection fixture. See docs/MODELS.md for the full results table and reproducer.

Tool-calling (used by agentic rescue) is reliable on Llama 3.1+, Qwen 2.5+/3+, DeepSeek-Coder-V2+, GPT-OSS, Gemma 3+, GLM 4+, Kimi K2+, and Granite 3. Smaller models (3B, 1B), thinking-token models (DeepSeek-R1 distills), and pre-3 Gemma fall back to patch-emit automatically.

If your hardware can't run a strong local model — or you want frontier-quality output for a tough adversarial review — Ollama Cloud exposes hosted models behind the same API via a :cloud suffix. The plugin treats them identically to local models; nothing in the plugin needs to change.

Cloud models send your diff context to Ollama's hosted endpoint — opt in by passing one explicitly via --model or /ollama:setup --default-model. Everything else stays local.

Override the model on any command with --model <name>.

Per-workspace config (set via /ollama:setup --default-model) is stored in the plugin state directory and takes precedence over OLLAMA_PLUGIN_DEFAULT_MODEL.

• Review and adversarial-review work on any model that produces valid JSON. Structured output uses Ollama's schema-constrained decoding (Ollama >= 0.5) for reliability.
• Rescue runs an agentic tool-calling loop by default: the model can read files, list directories, write files, apply patches, and run allowlisted commands autonomously (hard cap: 20 iterations). Use --emit-patch to force the legacy one-shot diff output instead. Models that do not support tool calling fall back to patch-emit automatically. Override the command allowlist with OLLAMA_PLUGIN_RESCUE_ALLOW_COMMANDS=cmd1,cmd2 (or =* for unrestricted).
• Stop-review gate uses a Stop hook — enable with /ollama:setup --enable-review-gate. It can create long Claude/Ollama loops; only enable when actively monitoring the session.
• Background jobs work for all long-running operations. Use --background and check progress with /ollama:status.
• Node.js 18.18 or later is required to run the companion script.

Beyond per-task dispatch, the plugin now supports a plan → execute → review pipeline that lets Claude delegate multi-step work to a chain of Ollama agents:

/ollama:plan "audit and fix error handling in lib/"     # planner reads the code, emits a structured plan
# Claude reviews. Looks good.
/ollama:execute-plan pln_abc                            # implement → verify → retry loop, autonomous per step
# Claude reviews the cumulative diff.

Three roles, three models (planner/implementer/verifier), Claude only at the gates. Typically 5–10× reduction in Claude tokens for refactors and audit-and-fix work. See docs/ORCHESTRATOR.md.

• docs/ORCHESTRATOR.md — orchestrator pipeline user guide
• docs/MODELS.md — empirical model recommendations + battle-test results
• docs/API.md — public vs internal surface for v1.x
• docs/SMOKE-TEST.md — pre-release verification checklist
• docs/PLAN-v1.md — v0.1 → v1.0 roadmap and status
• docs/PLAN-orchestrator.md — orchestrator design rationale
• CHANGELOG.md — release notes

Ported from openai/codex-plugin-cc, Apache 2.0. See NOTICE for attribution. This project is not affiliated with OpenAI or Anthropic.