Project planning with a local LLM — Ollama, no cloud required
Run shadow-planner's AI assistant and planning agent on a local model via Ollama. Draft plans from a brief and steer your Gantt chart by chat while nothing leaves your machine — or bring your own Claude, OpenAI or OpenRouter key.
"AI-assisted" usually means "sends your data to someone's API". shadow-planner is built the other way round: the default AI provider is a local model running on your own hardware through Ollama, and cloud models are an opt-in you control per request.
What the AI does
Two features share the same provider setting:
The planning agent turns a Markdown brief into a structured plan — epics, tasks with dates, dependencies, assignments — for you to review and materialise. Walkthrough →
The chat assistant is a conversational layer over your active scenario. It answers questions ("who's over capacity in March?"), makes changes ("move the beta out a week"), and narrates what a cascade will do before you approve it. It has a deliberately compact tool set — around 27 tools — so small local models stay reliable.
Running it locally
- Install Ollama (Windows, macOS, Linux).
ollama pull qwen3:8b— or any model you like; larger models draft better plans.- In shadow-planner's first-run wizard choose Ollama (local), pick the model, test the connection.
That's the whole setup. No key, no account, no per-token bill. shadow-planner includes a rescue parser for tool calls from smaller models, so you don't need a frontier model for everyday use.
When you want a bigger model
Switch providers in Settings at any time:
| Provider | What's sent | Notes |
|---|---|---|
| Ollama (default) | Nothing leaves the machine | Free, offline |
| Anthropic Claude | Your chat messages / brief, per request | Prompt caching keeps repeat calls cheap |
| OpenAI | Your chat messages / brief, per request | Chat Completions API |
| OpenRouter | Your chat messages / brief, per request | One key, hundreds of models from many vendors |
In every case the app talks to the provider directly from your machine with your own key. There's no shadow-planner server in between, and the database itself is never transmitted.
Guard rails regardless of model
- Approval gate. Deletes, bulk updates and date-shift cascades pause with a preview card. Decline and nothing is written.
- Audit log. Every write — by you or by the assistant — is recorded with who did it and what changed.
- Scoped to a scenario. The assistant only sees and touches the scenario you have open. Experiment in a sandbox scenario, keep the baseline clean.
- @mentions and slash commands. Pin the assistant to a specific
@projector@employee; use/statusand/exportfor deterministic, LLM-free actions.
Who this is for
Consultancies with client confidentiality clauses. Teams in regulated industries. Anyone who has been told "no external AI tools" but still wants the speed. And, honestly, anyone who'd rather not pay per token to iterate on a plan ten times.
The AI features are included in the free tier. Install shadow-planner.