Getting Started
Questi contenuti non sono ancora disponibili nella tua lingua.
Step 1: Install the prerequisites
Section titled “Step 1: Install the prerequisites”Soleri needs two things on your machine, plus one optional setup for native builds. If you already have these, skip to Step 2.
1. Node.js 18+
Section titled “1. Node.js 18+”Soleri runs on Node.js. Check if you have it:
node -vIf the command isn’t found or shows a version below 18:
| Platform | Install command |
|---|---|
| macOS | brew install node (requires Homebrew) or download from nodejs.org |
| Linux | curl -fsSL https://deb.nodesource.com/setup_22.x | sudo bash - && sudo apt-get install -y nodejs |
| Windows | Download the installer from nodejs.org |
npm ships with Node.js — no separate install needed.
2. An MCP-compatible AI editor
Section titled “2. An MCP-compatible AI editor”Soleri connects to your AI editor via MCP (Model Context Protocol). You need at least one of these installed:
| Editor | Install |
|---|---|
| Claude Code (recommended) | npm install -g @anthropic-ai/claude-code — see Claude Code docs |
| Codex | npm install -g @openai/codex — see Codex docs |
| OpenCode | go install github.com/opencode-ai/opencode@latest — see OpenCode docs |
3. Build tools (optional — for the Knowledge Engine)
Section titled “3. Build tools (optional — for the Knowledge Engine)”Soleri’s Knowledge Engine uses better-sqlite3, which requires native compilation. Scaffolding works without it, but the engine won’t start until build tools are available.
| Platform | Install command |
|---|---|
| macOS | xcode-select --install |
| Linux | sudo apt-get install -y build-essential python3 |
| Windows | Install Visual Studio Build Tools or use WSL |
Verify your setup
Section titled “Verify your setup”node -v # should print v18.x or highernpm -v # should print 8.x or higherclaude --version # if using Claude CodeStep 2: Create your agent
Section titled “Step 2: Create your agent”One command to scaffold a file-tree agent in your current directory:
npm create soleri my-agentThis creates ./my-agent/ in whatever directory you run it from.
The interactive wizard asks for:
| Prompt | What it means |
|---|---|
| Agent name | Your agent’s identity (e.g., “sentinel”, “architect”) |
| Role | One-line description of what it does |
| Domains | Knowledge areas: frontend, backend, security, or custom |
| Tone | How the agent communicates: precise, mentor, pragmatic |
This generates a folder — no TypeScript, no build step:
my-agent/├── agent.yaml # Identity + engine config├── agent-config.yaml # Capabilities, probes, workflow mappings├── .mcp.json # Connects to Knowledge Engine (Claude Code)├── opencode.json # Connects to Knowledge Engine (OpenCode)├── settings.local.json # Claude Code hooks + pre-approved permissions├── .gitignore # Excludes auto-generated files├── CLAUDE.md # Auto-generated (never edit)├── instructions/ # Behavioral rules│ ├── _engine.md # Engine rules (auto-generated)│ └── domain.md # Your domain-specific rules├── workflows/ # Step-by-step playbooks│ ├── feature-dev/│ ├── bug-fix/│ ├── code-review/│ └── context-handoff/├── flows/ # Flow YAML definitions├── knowledge/ # Domain intelligence bundles├── skills/ # SKILL.md files├── hooks/ # Your AI editor hooks├── data/ # Agent runtime data└── workspaces/ # Workspace contextsYour agent is ready to use immediately. No npm install, no npm run build.
Step 3: Register and start
Section titled “Step 3: Register and start”From inside the agent folder, register the MCP server and start:
cd my-agentsoleri install --target claude # Claude Code (default)soleri install --target opencode # OpenCodesoleri install --target codex # Codexsoleri install --target all # All editorssoleri dev # Start engine + watch for changessoleri install registers the Soleri Knowledge Engine in your editor’s MCP config. soleri dev starts the engine and watches your agent folder — CLAUDE.md is regenerated automatically when you edit agent.yaml or instructions/.
Step 4: Connect to your AI editor
Section titled “Step 4: Connect to your AI editor”After running soleri install, restart your AI editor. Your agent is available as a tool. The .mcp.json in your agent folder looks like:
{ "mcpServers": { "soleri-engine": { "command": "npx", "args": ["@soleri/engine", "--agent", "./agent.yaml"] } }}Step 5: First conversation
Section titled “Step 5: First conversation”Once connected, try these in your AI editor:
# Activate the persona"Hello, My Agent!"
# Search your agent's knowledge"Search for patterns about error handling"
# Capture something you learned"Capture this pattern: always use error boundaries at route level"
# Check agent health"Run a health check"Your agent starts with starter knowledge and learns from every session.
Updating soleri
Section titled “Updating soleri”# Check for updatessoleri agent status
# Update engine to latest versionnpx @soleri/cli agent update
# Regenerate CLAUDE.md with latest engine rulessoleri agent refreshTo re-scaffold from scratch (e.g., after a major version bump):
rm -rf ~/.npm/_npx # clear stale npx cachenpm create soleri@latest my-agentCustomize your agent
Section titled “Customize your agent”Edit files directly, no rebuild needed:
- Add rules by creating a new
.mdfile ininstructions/ - Add workflows by creating a new folder in
workflows/withprompt.md+gates.yaml - Add knowledge by dropping a JSON bundle in
knowledge/ - Change identity by editing
agent.yaml
Run soleri dev and CLAUDE.md regenerates automatically on save.
Health check
Section titled “Health check”If something isn’t working:
npx @soleri/cli doctorReports Node version, npm status, agent context, vault health, and engine connectivity.
What’s next
Section titled “What’s next”- Your first 10 minutes — a hands-on tutorial to see your agent in action
- The development workflow — the five-step rhythm: Search, Plan, Work, Capture, Complete
- CLI reference — every CLI command with options and examples
- Extending your agent — add instructions, workflows, knowledge, and packs
Need help?
Section titled “Need help?”If you run into issues, check Troubleshooting or reach out at hello@soleri.ai.