1. Install OpenPilot
Windows 10+ (x64) is available now. Grab the installer or portable build from the download page.
- Installer (.exe) — standard setup for most users
- Portable (.exe) — run without installing
macOS builds (Apple silicon & Intel) are coming soon.
2. Add a model
OpenPilot is a harness, not a model marketplace. You bring an OpenAI-compatible endpoint and key.
- Open Settings → Models
- Paste your base URL, API key, and model id
- Save and select the model in the composer
{
"name": "gpt-4.1",
"baseUrl": "https://api.openai.com/v1",
"apiKey": "sk-••••••••",
"modelId": "gpt-4.1"
}
Works with OpenAI, OpenRouter, Ollama (http://localhost:11434/v1), LM Studio, and custom gateways.
3. Open a workspace
Pick a project folder. The agent gets filesystem and shell access inside that workspace so it can create files, edit code, install deps, and run commands.
4. Run your first task
Describe the outcome in chat—build an API, fix a failing test, research a library, scaffold a README. Watch the tool trail as OpenPilot writes files, runs the terminal, and verifies results.
Example prompt:
Build a small Express API with a health check, auth middleware, and a README.
Live web research with TinyFish
When docs, APIs, or errors live outside the repo, OpenPilot can search and fetch pages in real time via TinyFish—then ground the answer in what it actually read.
- Open Settings → TinyFish (or the web research panel)
- Add your TinyFish API key so
web_searchandweb_fetchare available - Ask the agent to look something up—library docs, error messages, changelogs, or API references
web_search("node fetch timeout")
web_fetch(docs.example.com)
→ cited answer + fix
Without TinyFish configured, the agent still works for local filesystem and terminal tasks; web research tools stay unavailable until a key is set.
Skills & long-term memory
Drop reusable playbooks as skills. Persist project preferences in AGENTS.md. OpenPilot loads them into context and can update memory when you ask.
# AGENTS.md
- Prefer TypeScript strict
- Use pnpm in this repo
- Never commit .env
MCP servers
Connect Model Context Protocol servers in mcp.json, restart, then toggle them per chat when you need extra tools (docs, browsers, databases, internal APIs).
Tips
- Start with a clear outcome, then refine with follow-ups
- Review tool cards before approving sensitive shell commands
- Track token usage in-app—usage stays tied to your keys
- For local models, use a coding-capable model for multi-step work
- Configure TinyFish when the agent needs live docs or web research
Ready? Download OpenPilot · Read the FAQ