@silverbackbase/root

Business knowledge primitive for AI agents — a per-client "brain" in two layers:

• Soul — stable identity: offers, personas, objections, differentiators, competitors, geo zones, pricing, FAQs, profile. Extracted automatically at onboarding.
• Memory — accumulated experience: decisions and their reasons, patterns that repeatedly work or break, and never-do hard rules. Written as the work happens.

The soul tells an agent who the client is. The memory stops it from resurfacing a dead idea, breaking a constraint, or repeating last month's mistake.

How it works

Give Root a client URL. In 5 minutes, all your agents have full business context — no manual briefing.

URL → Firecrawl (site scraping) → LLM extraction → DataForSEO (local competitors) → structured knowledge

MCP Tools

| Tool | Type | Description | |------|------|-------------| | root_get_context | Read | Full business knowledge (soul + memory) — call before any content/SEO/prospection task. Optional task_type keeps it lean (filters decisions/patterns; soul + never-do always full) | | root_get_category | Read | Specific category (personas, objections, decisions, never-do, etc.) | | root_search | Read | Text search across all knowledge | | root_history | Read | Change timeline of an item (or recent project activity) — before/after snapshots | | root_onboard_from_url | Write | Onboard a client from their website URL | | root_refresh | Write | Re-query SERP and detect new/lost competitors — run weekly or on demand | | root_create_item | Write | Add a knowledge item (soul, or memory with required provenance; optional task_types to scope) | | root_update_item | Write | Update an existing item (provenance re-checked on memory items) | | root_delete_item | Write | Delete an item | | root_revert | Write | Roll an item back to a previous state (re-creates it if deleted) — never destructive |

Self-hosted

Requirements

• Node.js ≥ 22
• Firecrawl API key (for onboarding)
• Anthropic API key (for LLM extraction)
• DataForSEO credentials (for competitor analysis, optional)

No database setup required — Root uses SQLite by default.

Install

Add to your Claude Code MCP config (.mcp.json in your project, or ~/.claude/claude_desktop_config.json):

{
  "mcpServers": {
    "root": {
      "command": "npx",
      "args": ["-y", "--package=@silverbackbase/root", "root-mcp"],
      "env": {
        "ROOT_ACCOUNT_ID": "my-agency",
        "ANTHROPIC_API_KEY": "sk-ant-...",
        "FIRECRAWL_API_KEY": "fc-...",
        "DATAFORSEO_LOGIN": "...",
        "DATAFORSEO_PASSWORD": "..."
      }
    }
  }
}

Root creates a root.db SQLite file automatically on first run. No migration, no setup.

Cloud deployment

For multi-user or hosted setups, set DATABASE_URL to switch to PostgreSQL automatically:

DATABASE_URL=postgresql://user:password@host:5432/root
ROOT_ACCOUNT_ID=my-agency
ANTHROPIC_API_KEY=sk-ant-...
FIRECRAWL_API_KEY=fc-...

Root detects DATABASE_URL at startup and creates the table if it doesn't exist.

Environment variables

| Variable | Required | Description | |----------|----------|-------------| | ROOT_ACCOUNT_ID | No | Account identifier (default: "default") | | DB_PATH | No | SQLite file path (default: ./root.db). Ignored when DATABASE_URL is set. | | DATABASE_URL | No | PostgreSQL connection string. If set, Root uses Postgres instead of SQLite. | | ANTHROPIC_API_KEY | For onboarding | LLM extraction from website content | | FIRECRAWL_API_KEY | For onboarding | Website scraping | | DATAFORSEO_LOGIN | For onboarding + refresh | Competitor analysis via SERP | | DATAFORSEO_PASSWORD | For onboarding + refresh | Competitor analysis via SERP | | ROOT_REFRESH_CRON | No | Cron expression for automatic market refresh (e.g. 0 8 * * 1 = every Monday 8am). Scheduler is disabled if not set. |

Knowledge categories

Soul — stable identity

| Category | Contains | |----------|---------| | offer | Services, products, prices | | persona | Target customer profiles | | objection | Common sales objections and responses | | differentiator | Competitive advantages | | competitor | Competitor names, positioning, weaknesses | | geo | Geographic zones served | | pricing | Pricing structure and packages | | faq | Frequently asked questions | | profile | General business profile (name, sector, tone) |

Memory — accumulated experience

| Category | Contains | |----------|---------| | decision | A choice that was made, and the reasoning behind it | | pattern | Something that repeatedly works or breaks, by task type | | never_do | A hard constraint — what to never propose, write, or recommend |

Provenance is mandatory for memory. Every decision, pattern, and never_do item must include source (where the knowledge came from) and reason (why it matters) — otherwise the write is rejected. A brain is only useful if the team trusts it, and trust comes from knowing where each entry came from. No source, no entry.

// root_create_item — a never-do rule
{
  "category": "never_do",
  "data": {
    "rule": "Never position the client as the cheap alternative",
    "source": "client strategy call 2026-06-01",
    "reason": "Premium positioning — price-led messaging attracts the wrong leads"
  }
}

root_get_context returns memory alongside the soul, under constraints (never-do), decisions, and patterns. Treat constraints as non-negotiable.

Requirements

• Node.js 22+
• Claude Code or any MCP-compatible AI agent

License

MIT — SilverBackBase