examples

Aeon integration examples

Aeon ships an A2A gateway and an MCP server so any agent stack can call its 90+ skills. The scripts here are the shortest possible "first call works" demos — point them at a running gateway, run python <file>, get a real Aeon output back.

File	Stack	Skill called	What it shows
a2a/langchain_client.py	LangChain	aeon-fetch-tweets	Wrap a skill as a langchain.tools.Tool
a2a/autogen_workflow.py	AutoGen	aeon-deep-research	Function tool inside a multi-agent chat
a2a/crewai_task.py	CrewAI	aeon-pr-review	Subclass BaseTool, hand to a Crew agent
a2a/openai_agents_client.py	OpenAI Agents SDK	aeon-token-report	@function_tool decorator pattern
mcp/test_connection.py	MCP (stdio)	aeon-cost-report (default)	List + invoke any aeon-* tool
mcp/claude_desktop_config.json	Claude Desktop	—	Drop-in config snippet

Every A2A script is <100 lines, depends only on requests plus the framework SDK, and reads its endpoint from A2A_GATEWAY_URL (defaults to http://localhost:41241).

A2A — start the gateway, then run any client

# Terminal 1 — start the gateway from your aeon repo
./add-a2a                      # listens on http://localhost:41241

# Terminal 2 — point a client at it
export A2A_GATEWAY_URL=http://localhost:41241
pip install langchain requests
python examples/a2a/langchain_client.py "AI agents"

Swap the script for autogen_workflow.py, crewai_task.py, or openai_agents_client.py to drive the gateway from a different framework. The four files share the same submit/poll pattern so you can crib whichever one matches your stack.

To call a different skill, change the skillId (and the var if the skill needs one). The agent card at http://localhost:41241/.well-known/agent.json lists every tool, its description, and an example invocation.

MCP — verify the round-trip

./add-mcp --build-only          # produce apps/mcp-server/dist/index.js
pip install mcp                 # official Anthropic MCP client
python examples/mcp/test_connection.py

You should see the full list of aeon-* tools followed by a real aeon-cost-report output. Once that works, hand apps/mcp-server/dist/index.js to Claude Code with ./add-mcp (already done if you ran ./add-mcp without --build-only) or to Claude Desktop using mcp/claude_desktop_config.json — replace /ABSOLUTE/PATH/TO/aeon with your actual repo path.

Picking a different skill

skills.json at the repo root is the source of truth — every entry is callable as aeon-<slug> from MCP and as skillId: "aeon-<slug>" from A2A. Some good first calls:

• aeon-cost-report — fast, no external API needed, safe to run anywhere
• aeon-token-report (var=AEON) — public DexScreener data, no secrets required
• aeon-deep-research (var="your topic") — long-running; expect 5–10 min
• aeon-fetch-tweets (var="your topic") — needs XAI_API_KEY in the Aeon repo's environment

Skills that hit external APIs need the same secrets the Aeon GitHub Actions runner uses. Drop them into a .env file at the Aeon repo root before you start the gateway or MCP server.

What the gateway is doing under the hood

Every Aeon skill is a markdown prompt at skills/<slug>/SKILL.md. The A2A and MCP servers both spawn claude -p - with the same prompt the GitHub Actions runner uses — so a skill behaves identically whether it fires on a cron, from your terminal, or from a remote agent on the other side of the protocol. No re-implementation, no drift.