---

name: cloud-api-and-integrations description: "Routes PaddleOCR users to the hosted API, MCP server, and LangChain integration surfaces." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

Cloud API and Integrations

Use this route when the task depends on a hosted PaddleOCR service, an access token, a base URL, or an application integration such as MCP or LangChain.

Handle these tasks here

• Hosted OCR and document parsing through PaddleOCRClient and AsyncPaddleOCRClient.
• The paddleocr api CLI.
• paddleocr_mcp server configuration and provider selection.
• LangChain PaddleOCRVLLoader usage and metadata handling.
• API option objects, result saving, and hosted-service troubleshooting.
• Reference-only guidance for the TypeScript/Go API SDK docs under api_sdk/ when a user asks about the broader official API ecosystem.

Route away from here

• Local OCR, predictor classes, and engine/device tuning belong in local-ocr-pipelines.
• Full structured document pipelines and Office conversion belong in document-parsing-and-conversion.
• Training, export, deployment, and TIPC evidence belong in training-export-and-deployment.

Read these references

• references/official-api.md for the Python client, CLI payloads, job lifecycle, and result saving.
• references/mcp-and-langchain.md for MCP provider modes and the LangChain loader.
• references/troubleshooting.md for token, URL, timeout, provider, and resource issues.

Use the bundled script

• scripts/check_official_api_setup.py validates hosted-API setup without making a network request by default.

What future agents should know

• The hosted API client has a synchronous and an async form. Both use token-based auth and raise a dedicated PaddleOCR error hierarchy.
• The CLI and SDK distinguish OCR jobs from document-parsing jobs. Do not mix the wrong model_type, model, or options dataclass.
• MCP provider choice matters: local inference, AI Studio, Qianfan, and self-hosted modes have different env vars and model restrictions.
• The LangChain loader uses the hosted document-parsing API and returns Document objects with raw response metadata.
• Reference the integration docs before asking for a real remote call; a fake-client or payload-check test is often enough for diagnosis.

Common triggers

• "I have a PaddleOCR access token"
• "Use the hosted API instead of local inference"
• "Set up the MCP server"
• "Load a document through LangChain"
• "Why is the API request failing or timing out?"