Getting Started

Sign up at leftfold.io with email, Google, or GitHub. Your workspace is created automatically with:

• 7 internal agents — summarizer, entity extractor, image analyzer, embedder, structured data generator, linker, and kind detector. All active and ready.
• 11 aggregate types — Article, Image, Video, Audio, Document, Person, Place, Organization, Brand, Workspace, and Project. See Domain Model.
• Event storming — an optional onboarding step where you describe your business and LeftFold proposes custom aggregate types. See Custom Aggregates.

Add the following to your Claude Desktop config at ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "leftfold": {
      "url": "https://leftfold.io/mcp"
    }
  }
}

Restart Claude Desktop. On first use, it will walk you through OAuth sign-in:

1. Claude discovers LeftFold's OAuth metadata automatically
2. A browser window opens — sign in with your LeftFold account
3. Consent to the requested scopes (read and write access to your workspace)
4. Done — Claude stores the token and reconnects automatically

You'll see LeftFold's tools in the tool picker. Try whoami to verify — it returns your workspace name, email, and active scopes.

If your MCP client doesn't support OAuth (or you prefer explicit keys), create an API key from Settings → API Keys in the dashboard, then pass it as a header:

{
  "mcpServers": {
    "leftfold": {
      "url": "https://leftfold.io/mcp",
      "headers": {
        "Authorization": "Bearer sk_your_api_key"
      }
    }
  }
}

See API Keys & Scopes for details on key creation and scope selection.

In Claude Desktop, just say:

Create an article about ACL recovery timelines for athletes

Claude calls the aggregate.append tool, which appends an article.created event to the event store. Behind the scenes:

• The article's traits (content-bearing, publishable) trigger the matching agents
• The summarizer generates a summary, the entity extractor finds people and places, the embedder creates a vector for semantic search
• Enrichments that need review appear in your mailbox

Ask Claude to list your content:

List my articles

Claude calls aggregate.list with type: "article". The query engine folds the event stream into current state and returns the result.

You can also search semantically:

Search my content for rehabilitation exercises

This uses aggregate.search, which queries the knowledge graph via pgvector. Results are ranked by semantic similarity, not just keyword matching.

After a few moments, agents will have enriched your article. Ask Claude:

Show my pending mailbox tasks

You'll see tasks from the summarizer, entity extractor, structured data generator, and linker. Each contains the agent's proposed enrichment as an artifact.

Approve the ones you like:

Approve all pending mailbox tasks

Approved enrichments are folded into the article's projection. Next time you query the article, you'll see the summary, entities, and structured data included.