AI Agents¶

EcoSemantic provides specialized AI agents that automate complex LCA workflows. Agents combine a large language model with access to EcoSemantic's MCP tools, executing multi-step analyses autonomously.

How Agents Work¶

Each agent operates through a loop: it receives your request, decides which tools to call, executes them, interprets the results, and either calls more tools or returns a final answer. Agents maintain conversation history (24 hours) so you can ask follow-up questions.

You → Agent → LLM decides tool calls → MCP tools execute → LLM interprets → Response
         ↑                                                      ↓
         └──────────────── iterate if needed ───────────────────┘

Endpoint: https://agent.ecosemantic.com/mcp

Tier requirement: Professional ($84.99/mo)

Authentication: Same OAuth 2.1 flow as the main MCP server. Agent tool calls count against your subscription.

Preset Agents¶

Five ready-to-use agents cover the core LCA workflow:

Agent	Role	Max Iterations	Best For
SCOUT	Discovery	8	Finding activities, methods, biosphere flows
ANALYST	Calculation	12	Carbon footprints, impact assessments, comparisons
TRACER	Supply Chain	8	Mapping suppliers, geographic dependencies
MODELER	Custom LCA	10	Building product models not in Ecoinvent
REPORTER	Synthesis	6	Executive summaries, comparison reports

SCOUT — Discovery Agent¶

Finds and recommends the right Ecoinvent data for your question. Searches activities with multiple keyword variations, filters by location, and returns activity codes ready for calculation.

Example:

"Find all lithium-ion battery production activities in Asia"

ANALYST — Calculation Agent¶

Executes LCA calculations end-to-end. Searches for the right impact method, runs calculate_standard or calculate_custom, polls for results, and interprets them with scientific context.

Example:

"Calculate the carbon footprint of 1 kWh of solar thermal electricity 
 in RoW using ecoinvent 3.12. Activity code: be60bf30bc2f744821eed45b46c6329e"

TRACER — Supply Chain Agent¶

Uses Neo4j graph queries to trace multi-tier supply chains, identify geographic dependencies, and find critical material hotspots. Executes Cypher queries against the Ecoinvent knowledge graph (26,533 activities, 9,850 biosphere flows, 335 locations).

Example:

"Trace the tier-1 and tier-2 suppliers for battery cell production in China"

MODELER — Custom LCA Agent¶

Builds custom product models by creating projects, activities, and exchanges linked to Ecoinvent background data. Follows ISO 14040/14044 conventions: each activity produces exactly 1 unit of output.

Example:

"Create a cradle-to-gate model for a 250g yogurt cup with PP packaging"

REPORTER — Synthesis Agent¶

Synthesizes calculation results, comparison data, and supply chain findings into clear reports and actionable recommendations. Works best after other agents have generated data.

Example:

"Summarize the environmental profile of steel production: 
 electric arc furnace vs blast furnace"

Agent Collaboration¶

Agents can be chained together for complex analyses:

SCOUT finds the right activities and their codes
ANALYST calculates environmental impacts
TRACER maps supply chain dependencies
MODELER builds custom models for products not in Ecoinvent
REPORTER synthesizes everything into a report

Each agent returns a conversation_id that you can use to continue the conversation with follow-up questions.

Conversation Persistence¶

All agents maintain conversation history stored in Redis with a 24-hour TTL. Pass the conversation_id from a previous response to continue:

{
  "message": "Now compare that with wind power",
  "conversation_id": "7f3562ad-2575-4c26-a34e-a1e2a7d2471f"
}

Custom Agents¶

Beyond the preset agents, you can create your own specialized agents with:

Custom system prompts — Define exactly how the agent should behave
Model selection — Choose from multiple LLMs via OpenRouter
MCP server selection — Pick which tool servers the agent can access
Configurable iterations — Set how many tool calls the agent can make (1-20)

Custom agents expire after 24 hours.

Available Models¶

Model	Provider	Cost Tier	Notes
`google/gemini-3-flash-preview`	Google	Free	Default — fast, no cost
`x-ai/grok-4.1-fast`	xAI	$	Fast with automatic caching
`nvidia/nemotron-3-nano-30b`	NVIDIA	$	Cost-effective
`zhipu/glm-4.7`	Zhipu	$$$	Strong multilingual support
`moonshotai/kimi-k2`	Moonshot	$$$$	Advanced reasoning

Available MCP Servers¶

Server ID	Description
`ecosemantic-lca`	Main LCA tools — Ecoinvent search, calculations, graph queries, custom modeling
`sap-mock`	SAP S/4HANA mock — product master data, BOMs, supplier locations

Workflow¶

1. list_available_models()           → See model options
2. list_available_mcp_servers()      → See server options  
3. create_custom_agent(              → Create the agent
     name="Battery Expert",
     system_prompt="You specialize in battery LCA...",
     model="google/gemini-3-flash-preview",
     mcp_servers=["ecosemantic-lca"]
   )
4. chat_with_custom_agent(           → Use it
     agent_id="...",
     message="Compare NMC vs LFP battery chemistries"
   )
5. delete_custom_agent(agent_id)     → Clean up when done

Management Tools¶

Tool	Description
`list_agents`	List all preset and custom agents
`list_available_models`	Available LLM models for custom agents
`list_available_mcp_servers`	Available MCP servers for custom agents
`create_custom_agent`	Create a new custom agent
`chat_with_custom_agent`	Send messages to a custom agent
`get_custom_agent`	View agent configuration
`list_custom_agents`	List your active custom agents
`delete_custom_agent`	Delete a custom agent

Technical Details¶

LLM Backend: OpenRouter API (supports Anthropic prompt caching for cost savings)
Tool Access: Agents call EcoSemantic MCP tools using the same OAuth token as the user
Rate Limiting: Agent tool calls count against user subscription quotas
Storage: Conversations and custom agent configs stored in Redis (DB 3)
Timeout: 300s per LLM call, conversation history TTL 24 hours
Port: 8884 (behind nginx at agent.ecosemantic.com)