Knowledge should compound. Here's how.

Vault, brain, and transport layers run on a single engine. One process, zero lock-in — swap transports, plug in knowledge, switch providers.

Agent

Your agent's config — agent.yaml defines identity, voice, domain wiring. Create in one command, customize anytime.

Domains

Pluggable knowledge domains. The engine loads relevant domains based on your agent's vault structure.

Engine

Single MCP server process. Vault (knowledge), Brain (learning), Memory (context), Planning (workflows). Your agent runs on it. Updates via npm update.

Transports

Protocol adapters isolated from core: mcp.ts ships now, rest.ts and lsp.ts plug in for VS Code, Cursor, Zed, and custom dashboards. Adding a transport = one adapter file.

Knowledge that compounds

Without a vault, every session starts from scratch. Soleri stores every pattern, anti-pattern, and decision in a searchable vault. Ask a question — get your team's answer, not a generic one.

Patterns, anti-patterns, and session memories — all searchable.

# Vault entry structure id: pattern-semantic-tokens type: pattern content: Always use semantic tokens context: domain: design-systems confidence: 0.94 sessions: 12 tags: [tokens, css, tailwind]

# Search your vault $ soleri vault search "button styling" Found 3 matches: pattern-semantic-tokens 94% confidence pattern-button-sizes 87% confidence anti-pattern-inline 91% confidence

# Brain strengths output $ soleri brain strengths Top patterns (last 7 days): semantic-tokens strength: 94 12 sessions component-structure strength: 87 8 sessions error-boundaries strength: 82 6 sessions a11y-focus-rings strength: 78 5 sessions zustand-patterns strength: 71 4 sessions

semantic-tokens

components

error-boundaries

a11y-focus

zustand

Gets sharper with every session

Manual knowledge curation doesn't scale. The brain tracks which patterns you apply, how often, and how successfully. Confidence grows automatically — no manual tagging needed.

Automatic capture, no manual tagging.

Knowledge that grows — and stays sharp

Patterns flow through a four-step lifecycle. The engine watches your work, captures what sticks, and compounds it over time.

Capture Engine notices repeated corrections, suggests patterns

↓

Store Patterns, anti-patterns, session memories enter the vault

↓

Strengthen Confidence scores track how often patterns succeed

↓

Compound High-confidence patterns surface first, weak ones flagged

Your vault doesn't decay — it sharpens. The brain tracks what works.

# Session 3 — engine notices a pattern ! You've corrected bg-blue-500 → bg-primary 3 times. Capture as pattern? [y/n] $ y ✓ Captured: semantic-token-enforcement ✓ Confidence: 0.85 (will increase with use)

# Vault entry with lifecycle metadata id: pattern-semantic-token-enforcement type: pattern content: Replace raw Tailwind colors with semantic tokens confidence: 0.91 # strengthened over 8 sessions sessions: 8 last_applied: 2 hours ago status: compounding # surfaces first in search

Not locked to any editor

Your knowledge shouldn't be trapped in one tool. The core engine has zero protocol dependencies. Transport adapters are separate files — swap or add them without touching the engine.

Core has zero protocol deps.

engine/transports/ # one file per protocol ├── mcp.ts # Claude Code, Cursor — ships now ├── rest.ts # dashboards, custom APIs └── lsp.ts # VS Code, Cursor, Zed extensions engine/core/ # zero transport imports ├── vault.ts ├── brain.ts └── memory.ts

# agent.yaml — LLM config llm: provider: anthropic # or: openai, ollama, none model: claude-sonnet-4-6 fallback: none # works without API keys

Anthropic OpenAI Ollama None

Works without API keys

Knowledge infrastructure shouldn't require a subscription. Vault search, pattern matching, and brain tracking work locally. LLM provider is optional — add one when you want AI-powered suggestions.

Basic functionality guaranteed without keys.

See what compounding knowledge looks like.

Install Soleri and watch the vault grow with every session.

Get started Your Agent