agentic-workflows

The same hello-agent, built once per designer, run across six agentic workflow tools: n8n, Flowise, Dify, Trigger.dev, LangGraph (MIT core), CrewAI. All six are self-hosted on one machine, all six call the same local model through a LiteLLM router, and every lane returned the same proof token. Zero cloud API keys were involved in any of the runs.

Why a repo of hello-worlds matters: the point is not the agent, the point is the comparison. Every designer is deployed from the compose files in its directory, every run is witnessed by the checked-in JSON in results/, and the whole estate is driven through one 235-line MCP server so an AI assistant operates all six through one surface (mcp/agentic_mcp.py, keys read from the OS keychain, values never printed).

Verify the claim without deploying anything

This is the whole smoke:

grep -h "HELLO-AGENT" results/*.json   # six lanes, six tokens

results/n8n-run.json records the webhook kick and the execution status instead of an echoed token (n8n's webhook returns "Workflow was started", and the execution log holds the model output).

Layout

Directory	What it is
designers/n8n/	n8n 3.0.1 compose; workflow Hello Agent RIG proof (exported in the run proof)
designers/flowise/	Flowise 3.1.3 compose; Agentflow v2 proof flow
designers/dify/	Dify 1.17.1; workflow + advanced-chat apps defined in-console, run proofs checked in
designers/trigger/	Trigger.dev self-hosted webapp compose (+ ClickHouse, Postgres, Redis, MinIO)
designers/langgraph/	LangGraph MIT core server: the agent graph (app/agent.py) and its serve config
designers/crewai/	CrewAI FastAPI service: the crew (app/crew.py) and its server
mcp/agentic_mcp.py	one stdio MCP server, six designers, scoped keychain reads
scripts/	env generation (keychain → compose env file) and deploy orchestration
results/	the run evidence, HTTP-level, per designer

Standing up your own

./scripts/deploy.sh            # compose up the six designers (docker required)
./scripts/generate-env.sh      # build the env file from OS keychain entries

Point the LiteLLM router (or any OpenAI-compatible endpoint) at a model, set OPENAI_BASE_URL, and the same hello-agent runs in each lane.

Honest limits

• The five proof tokens were produced against one local model (long-context-local alias); quality of the replies reflects that model, not the designers.
• n8n and Flowise store their definitions in their own databases; the directories here hold the deployment, and the proofs hold the flow identifiers. Full node-level exports per flow are in the works.
• LangGraph Studio and LangSmith hosted need licenses; this farm runs the MIT core server only.

Part of the mrodgersjs-web FDE studio fleet — control planes, gates, fleets, proof. rodgersintelligence.com