Simple Alfresco Agent Mesh

A hands-on lab to compare three routing options (cagent, LangGraph, LlamaIndex) over the same Alfresco setup

The mcp-client always talks to MCP Server available in localhost:8085; one router is enabled at a time, which then fans out to MCP Audit and MCP Docs, both talking to the Alfresco Repository

The stack that spins up:

• Alfresco Community Repository including Audit sample configuration
• Two MCP servers:
  • Audit MCP
    • Spring AI, streamable HTTP (exposes audit-centric tools)
    • Source code available in https://github.com/jottley/alfresco-mcp
  • Docs MCP
    • Python alfresco-mcp server (content/doc tools)
    • Source code available in https://github.com/stevereiner/python-alfresco-mcp-server
• An MCP Router/Orchestrator (cagent, LangGraph or LlamaIndex): service that classifies the user prompt and forwards the MCP call to the right backend (audit vs docs)

Designed to be easy to run locally, easy to understand, and a good template for building more sophisticated Agent Meshes

What’s inside

.
├── agent-mesh-tools
│   ├── cagent/               # Router with Docker cagent (LLM required)
│   ├── langgraph/            # Router with LangGraph (deterministic, no LLM)
│   └── llamaindex/           # Router with LlamaIndex (LLM required)
├── alfresco-agents
│   ├── alfresco/             # Alfresco Community stack (Repo, Share, Search, etc.)
│   ├── alfresco-mcp-audit/   # MCP Audit server
│   ├── alfresco-mcp-server/  # MCP Docs server
│   └── compose.yaml          # Always-on: Repo + MCP backends
├── mcp-client/               # Client used to test the active router
└── compose.yaml              # Root compose with `include:` and profiles per router

Architecture

flowchart LR
  %% Client
  C[mcp-client]

  %% Routers
  subgraph Routers
    CA[cagent - local app]
    LG[langgraph - router:8085]
    LI[llamaindex - router:8085]
  end

  %% MCP backends
  subgraph MCP["MCP Servers"]
    AU[MCP Audit - 8081]
    DO[MCP Docs - 8003]
  end

  %% Alfresco core
  subgraph Alfresco
    RE[Alfresco Repository - 8080]
  end

  %% Connections
  C --> CA
  C --> LG
  C --> LI

  CA --> AU
  CA --> DO
  LG --> AU
  LG --> DO
  LI --> AU
  LI --> DO

  AU --> RE
  DO --> RE

Loading

• The Router is the single MCP endpoint for clients
• It classifies each request and delegates to either the Audit MCP or Docs MCP backend over Streamable HTTP
• Both backends talk to Alfresco

Routers at a glance (pick one)

• LangGraph — deterministic router, no LLM required. Great for repeatable flows and simple, auditable routing rules.
• LlamaIndex — LLM-driven tool routing. Uses the model to choose between Audit/Docs tools; more adaptive, needs an LLM.
• cagent — LLM-driven agent runtime (YAML) with first-class MCP tool integration. It runs standalone as a CLI/TUI agent, as described in Instantly Build AI Agents for Alfresco with Docker new "cagent"

Requirements

• Docker & Docker Compose v2+ Needed to run the full stack (Alfresco, MCP servers, routers, client).

• System resources At least 8 GB RAM free and 4 CPU cores recommended.

• Open ports

  • 8080 > Alfresco Repository
  • 8081 > MCP Audit
  • 8003 > MCP Docs
  • 8085 > Active router (LangGraph, or LlamaIndex)

• LLM runtime

  • For cagent and LlamaIndex, an LLM is required
  • This lab uses Ollama with the local model gpt-oss
  • Ensure Ollama is installed and the model is pulled before starting:

    ollama pull gpt-oss

• No LLM required

  • LangGraph router works deterministically without an LLM

Quick start

1. Start Alfresco Community + MCP backends

docker compose up --build

2. Activate exactly one router (both expose :8085):

docker compose --profile langgraph up -d

docker compose --profile llamaindex up -d

For cagent follow the steps in Instantly Build AI Agents for Alfresco with Docker new "cagent"

3. Test with the client (always hits 8085)

docker compose -f mcp-client/compose.yaml run --rm mcp-client

4. Switch routers (keep Alfresco running)

docker compose --profile langgraph down
docker compose --profile llamaindex up --build

Example prompts (copy/paste)

[DOCS]

> I need to search for documents in Alfresco. Can you search for:
 * Documents containing "budget"
 * Maximum 5 results

[AUDIT]

> List Alfresco audit apps

> List Alfresco audit entries for audit app "search"

Tips:

• If your router is LangGraph, routing is rule-based—logs/retention/compliance > Audit; search/metadata/renditions/transforms > Docs
• If it’s LlamaIndex or cagent, ensure a model is configured (e.g., OpenAI/Anthropic/Gemini or a local model via Docker Model Runner)

Lab flow (recommended)

1. Bring up Alfresco + MCP (wait for health)
2. Start LangGraph first (no keys needed) and run the prompts above
3. Stop LangGraph, start LlamaIndex; repeat the prompts and observe differences in routing/answers
4. Stop LlamaIndex, follow the steps in Instantly Build AI Agents for Alfresco with Docker new "cagent"

License

This repository is provided under the Apache 2.0 license (see LICENSE). Upstream components maintain their own licenses.