RAGfly — MCP Interface

Connect any MCP-compatible LLM agent (Claude Code, Cursor, Cline, etc.) to your RAGfly group's documents and capabilities. The agent discovers tools automatically — no integration code required.

Prerequisite

An API Key for your group. See INTEGRATION.md § Credentials.

Quick setup (remote mode — recommended)

No installation required. Add to your MCP client:

Claude Code — .mcp.json (project) or ~/.mcp.json (global)

SSE (compatible with all clients):

{
  "mcpServers": {
    "ragfly": {
      "url": "https://api.ragfly.ai/mcp/sse",
      "headers": {
        "Authorization": "Bearer slm_live_xxxxxxxxxx"
      }
    }
  }
}

streamable_http (more efficient, better with HTTP/2 and proxies):

{
  "mcpServers": {
    "ragfly": {
      "url": "https://api.ragfly.ai/mcp-http/",
      "headers": {
        "Authorization": "Bearer slm_live_xxxxxxxxxx"
      }
    }
  }
}

Restart your client. Tools appear with the prefix mcp__ragfly__.

Cursor / Cline / other MCP clients

Register the SSE URL and the Authorization header. Consult your client's documentation for the exact format — the protocol is standard MCP.

Available tools

Always call estado_sesion first to confirm the connection is valid.

Document estado values

CARGADO · METADATA · ESCANEADO · CHUNKEADO · VECTORIZADO · NO_ESCANEABLE · REVISAR

Queue estado values

PENDIENTE · EJECUTANDO · COMPLETADO · ERROR

Example flow (agent)

# 1. Verify connection
estado_sesion()
→ {"usuario": "bot-finance", "grupo": "COMPANY", "rol": "DOC-ADMIN"}

# 2. Ask over documents
preguntar(texto="What are the penalty clauses in the 2024 contracts?")
→ RAG response with chunk citations

# 3. List vectorized documents
listar_documentos(estado="VECTORIZADO", limite=10)

# 4. Execute a skill over a workspace
ejecutar_habilidad(codigo_habilidad="RESUMIR_DOCUMENTO", id_espacio=42)

# 5. Monitor progress
ver_cola(estado="EJECUTANDO")

Local stdio mode (alternative)

If you already have the RAGfly client installed and need to operate on the local filesystem:

pip install ragfly
ragfly login

Config in .claude/settings.json:

{
  "mcpServers": {
    "ragfly-local": {
      "command": "ragfly-mcp",
      "args": []
    }
  }
}

The local mode exposes the same tools but reads the JWT from ~/.ragfly/credentials.json. Useful when the agent also needs to scan local directories with ragfly local escanear.

Permissions

Each tool operates in the context of the API Key's user — same RBAC as the web interface. An agent with role DOCS-USUARIO-FINAL can read but cannot execute skills.

Troubleshooting

Codex

Codex supports MCP via codex mcp add. Use the streamable_http endpoint:

codex mcp add ragfly \
  --url https://api.ragfly.ai/mcp-http/ \
  --bearer-token-env-var RAGFLY_API_KEY

See QUICKSTART.md for the full walkthrough with setup, test script, and case table.

Practical differences