Part 4 — Integrations¶
Multimodal processing and LangGraph interoperability.
4.1 Multimodal pipeline (Example 15)¶
Process images, audio, and video in a unified pipeline (mock providers in the example; swap for real APIs in production).
python examples/15_multimodal_pipeline.py
Typical outputs:
| Modality | Fields |
|---|---|
| Image | description, labels, confidence |
| Audio | text (transcript), language |
| Video | summary, key_frames, scenes |
Wire real providers by implementing the multimodal adapter interfaces in hiveflow.multimodal.
4.2 LangGraph export (Example 16)¶
Export a HiveFlow TaskGraph to LangGraph-compatible JSON and Python scaffold.
python examples/16_langgraph_export.py
The export includes:
state_schema(messages, results, intent_id)nodeswith actions/skillsedgesincluding__start__/__end__- Optional
interrupt_beforefor HITL nodes
Generated Python uses langgraph + langchain-core (PoC — see versioning policy).
Sidecar pattern¶
Keep LangGraph as the execution runtime and HiveFlow as the coordination layer (HITL, audit, blackboard, Studio UI):
User → Studio → HiveFlow (HITL + audit + blackboard)
↓
LangGraph runtime (graph execution)
Deep dive: LangGraph Sidecar cookbook and Ecosystem compatibility guide.
Studio export buttons¶
In Orchestrator (Agent mode):
- plan-only → Export LangGraph JSON or Export LangGraph + Python template
- Toolbar Export LangGraph — converts current canvas via
POST /api/agent/export-langgraph
4.3 LLM provider integrations¶
| Provider | Doc | Env vars |
|---|---|---|
| OpenAI | integrations/openai.md | OPENAI_API_KEY |
| Anthropic | integrations/anthropic.md | ANTHROPIC_API_KEY |
| Ollama | Agent llm/ollama_client.py |
OLLAMA_BASE_URL |
| DeepSeek | Agent llm/deepseek_client.py |
DEEPSEEK_API_KEY |
Route planning vs execution separately:
HIVEFLOW_LLM_PLANNING_PROVIDER=openai
HIVEFLOW_LLM_EXECUTION_PROVIDER=anthropic
4.4 MCP ecosystem¶
HiveFlow speaks MCP natively (hiveflow.mcp). Plugins register tools discoverable by any MCP-compatible agent. See Part 2 — MCP.
4.5 Exercises¶
- Export Example 02's pipeline to LangGraph JSON and inspect node IDs.
- Replace mock image analysis in Example 15 with a real vision API call.
- Run LangGraph sidecar with
HIVEFLOW_PLAN_HITL=trueand approve a plan in Studio.
Next¶
→ Part 5 — Studio: full visual ops UI walkthrough.