Agents

Agents built with Go Micro.

• Alert: record incidents and deliver failure and recovery notifications.
• DevOps: investigate service failures using health observations and logs.
• Monitor: check that the agents respond and record health changes.
• Release: release and deploy through a configured GitHub Actions workflow.
• Support: investigate GitHub issues and draft responses with supporting evidence.

Each agent is a Go program. Service integrations live in services/ and can also be called directly over RPC. Run agents in your own environment with your own model configuration.

By default, Support covers this repository and Go Micro. Monitor, Alert and DevOps look after the agents running here. Override their targets for your own environment; see deploy for server installation.

Install

Requires Go 1.26 or newer, with Go's bin directory on your PATH.

go install go-micro.dev/v6/cmd/micro@master
git clone https://github.com/micro/agents
cd agents

Set your model credentials. For example, with OpenAI:

export MICRO_AI_PROVIDER=openai
export MICRO_AI_API_KEY=your-model-key
micro run

In another terminal, with the same model settings:

micro chat

micro run uses micro.json to build and start all five agents; Ctrl+C stops them. Release starts in inspection-only mode until you configure a workflow. Monitor and Alert work in the background without model calls. Each agent has its own persistent store database. Run chat on the same host with the same registry.

MCP

Expose running agents and services to an MCP client through Go Micro's gateway. With micro run running, start the gateway in another terminal on the same host:

micro gateway --mcp-address 127.0.0.1:3000

Add a Streamable HTTP MCP server in your client with this URL:

http://127.0.0.1:3000/mcp

The gateway automatically exposes discovered service endpoints as tools. The client lists and calls them without configuring each tool individually. The gateway and agents must use the same registry; the default is local mDNS. An MCP client only needs the gateway connection, not registry access.

For remote access, put the gateway behind HTTPS, enable --auth with MICRO_AUTH_TOKEN, and configure the client to send that bearer token. Use the public HTTPS URL ending in /mcp, rather than the loopback URL above.

To expose only a subset, optionally repeat --mcp-tool <tool-name>. The default includes write endpoints such as Release's dispatch operation. MCP calls do not inherit chat approval; restrict exposed tools before giving clients access to deployment actions.

Development

go test -race ./...
go build ./...

To test the full CLI journey with local model/webhook fixtures and a NATS registry, install nats-server, then run python3 scripts/smoke.py. Set MICRO_BIN to select a CLI binary. The test starts all agents, delegates from one chat, reads failure logs through DevOps, stops and restarts Support, verifies failure/recovery delivery, and checks Alert persistence. It uses temporary state and no real provider credentials.

Licensed under AGPL-3.0; see LICENSE.