station

Station - AI Agent Orchestration Platform

Build, test, and deploy intelligent agent teams. Self-hosted. Git-backed. Production-ready.

Quick Start | Real Example | Deploy | Documentation

---

Why Station?

Build multi-agent systems that coordinate like real teams. Test with realistic scenarios. Deploy on your infrastructure.

Station gives you:

• ✅ Multi-Agent Teams - Coordinate specialist agents under orchestrators
• ✅ Built-in Evaluation - LLM-as-judge tests every agent automatically
• ✅ Git-Backed Workflow - Version control agents like code
• ✅ One-Command Deploy - Push to production with stn deploy
• ✅ Full Observability - Jaeger traces for every execution
• ✅ Self-Hosted - Your data, your infrastructure, your control

---

Quick Start (2 minutes)

Prerequisites

• Docker - Required for Jaeger (traces and observability)
• AI Provider - Choose one:
  • CloudShip AI (Recommended) - STN_CLOUDSHIP_KEY or CLOUDSHIPAI_REGISTRATION_KEY
  • OPENAI_API_KEY - OpenAI (gpt-5-mini, gpt-5, etc.)
  • GEMINI_API_KEY - Google Gemini
  • ANTHROPIC_API_KEY - Anthropic (claude-sonnet-4-20250514, etc.)

1. Install Station

curl -fsSL https://raw.githubusercontent.com/cloudshipai/station/main/install.sh | bash

2. Initialize Station

Choose your AI provider:

Use CloudShip AI for optimized inference with Llama and Qwen models. This is the default when a registration key is available.

# Set your CloudShip registration key
export CLOUDSHIPAI_REGISTRATION_KEY="csk-..."
# Or use: export STN_CLOUDSHIP_KEY="csk-..."

stn init --provider cloudshipai --ship  # defaults to cloudship/llama-3.1-70b

Available models:

• cloudship/llama-3.1-70b (default) - Best balance of performance and cost
• cloudship/llama-3.1-8b - Faster, lower cost
• cloudship/qwen-72b - Alternative large model

⚠️ DEPRECATED: Anthropic OAuth is currently unavailable.

Anthropic has restricted third-party use of OAuth tokens. This authentication method is not working until further notice.

Please use one of the following alternatives:

• OpenAI API Key (recommended)
• Google Gemini API Key
• Anthropic API Key (pay-per-token, not subscription-based)

Use your existing Claude Max or Claude Pro subscription - no API billing required.

# ❌ NOT WORKING - Anthropic OAuth disabled
# stn init --provider anthropic --ship
# stn auth anthropic login

export OPENAI_API_KEY="sk-..."
stn init --provider openai --ship  # defaults to gpt-5-mini

export GEMINI_API_KEY="..."
stn init --provider gemini --ship

This sets up:

• ✅ Your chosen AI provider
• ✅ Ship CLI for filesystem MCP tools
• ✅ Configuration at ~/.config/station/config.yaml

3. Start Jaeger (Tracing)

Start the Jaeger tracing backend for observability:

stn jaeger up

This starts Jaeger UI at http://localhost:16686 for viewing agent execution traces.

4. Connect Your MCP Client

Choose your editor and add Station:

claude mcp add station -e OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318 --scope user -- stn stdio

Verify with claude mcp list.

Add to opencode.jsonc:

{
  "mcp": {
    "station": {
      "enabled": true,
      "type": "local",
      "command": ["stn", "stdio"],
      "environment": {
        "OTEL_EXPORTER_OTLP_ENDPOINT": "http://localhost:4318"
      }
    }
  }
}

Add to .cursor/mcp.json in your project (or ~/.cursor/mcp.json for global):

{
  "mcpServers": {
    "station": {
      "command": "stn",
      "args": ["stdio"],
      "env": {
        "OTEL_EXPORTER_OTLP_ENDPOINT": "http://localhost:4318"
      }
    }
  }
}

OS	Config Path
macOS	~/Library/Application Support/Claude/claude_desktop_config.json
Windows	%APPDATA%\Claude\claude_desktop_config.json
Linux	~/.config/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "station": {
      "command": "stn",
      "args": ["stdio"],
      "env": {
        "OTEL_EXPORTER_OTLP_ENDPOINT": "http://localhost:4318"
      }
    }
  }
}

Optional GitOps: Point to a Git-backed workspace:

"command": ["stn", "--config", "/path/to/my-agents/config.yaml", "stdio"]

5. Install Editor Plugins (Optional)

Get skills, slash commands, and enhanced documentation for your AI editor:

Adds /station commands, skills for agent creation, and MCP server config.

# Add Station marketplace and install plugin
/plugin marketplace add cloudshipai/station
/plugin install station@cloudshipai-station

Or install from local clone:

/plugin install ./station/claude-code-plugin

Adds Station CLI reference skill with agent, workflow, and deployment docs.

# Copy skill to your project
cp -r station/opencode-plugin/.opencode .

# Or install globally
cp -r station/opencode-plugin/.opencode ~/.config/opencode/

Restart OpenCode - skill auto-loads.

6. Start Building

Restart your editor. Station provides:

• ✅ Web UI at http://localhost:8585 for configuration
• ✅ Jaeger UI at http://localhost:16686 for traces
• ✅ 41 MCP tools available in your AI assistant

Try your first command:

"Show me all Station MCP tools available"

Copy this prompt into your AI assistant for a hands-on tour:

You are my Station onboarding guide. Walk me through an interactive hands-on tutorial.

RULES:
1. Create a todo list to track progress through each section
2. At each section, STOP and let me engage before continuing
3. Use Station MCP tools to demonstrate - don't just explain, DO IT
4. Keep it fun and celebrate wins!

THE JOURNEY:

## 1. Hello World Agent
- Create a "hello-world" agent that greets users and tells a joke
- Call the agent and show the result
[STOP for me to try it]

## 2. Faker Tools & MCP Templates
- Explain Faker tools (AI-generated mock data for safe development)
- Note: Real MCP tools are added via Station UI or template.json
- Explain MCP templates - they keep credentials safe when deploying
- Create a "prometheus-metrics" faker for realistic metrics
[STOP to see the faker]

## 3. DevOps Investigation Agent
- Create a "metrics-investigator" agent using our prometheus faker
- Call it: "Check for performance issues in the last hour"
[STOP to review the investigation]

## 4. Multi-Agent Hierarchy
- Create an "incident-coordinator" that delegates to:
  - metrics-investigator (existing)
  - logs-investigator (new - create a logs faker)
- Show hierarchy structure in the .prompt file
- Call coordinator: "Investigate why the API is slow"
[STOP to see delegation]

## 5. Inspecting Runs
- Use inspect_run to show detailed execution
- Explain: tool calls, delegations, timing
[STOP to explore]

## 6. Workflow with Human-in-the-Loop
- Create a workflow: investigate → switch on severity → human_approval if high → report
- Make it complex (switch/parallel), not sequential
- Start the workflow
[STOP for me to approve/reject]

## 7. Evaluation & Reporting
- Run evals with evaluate_benchmark
- Generate a performance report
[STOP to review]

## 8. Grand Finale
- Direct me to http://localhost:8585 (Station UI)
- Quick tour: Agents, MCP servers, Runs, Workflows
- Celebrate!

## 9. Want More? (Optional)
Briefly explain these advanced features (no demo needed):
- **Schedules**: Cron-based agent scheduling
- **Sandboxes**: Isolated code execution (Python/Node/Bash)
- **Notify Webhooks**: Send alerts to Slack, ntfy, Discord
- **Bundles**: Package and share agent teams
- **Deploy**: `stn deploy` to Fly.io, Docker, K8s
- **CloudShip**: Centralized management and team OAuth

Start now with Section 1!

---

Running Station with stn up

The easiest way to run Station is with stn up - a single command that starts Station in a Docker container with everything configured.

Primary Use Case: Running Bundles

stn up is designed to make it trivial to run agent bundles from your CloudShip account or the community:

# Run a bundle from CloudShip (by ID or name)
stn up --bundle finops-cost-analyzer

# Run a bundle from URL
stn up --bundle https://example.com/my-bundle.tar.gz

# Run a local bundle file
stn up --bundle ./my-custom-agents.tar.gz

This is the recommended way for most users to get started - just pick a bundle and go.

Secondary Use Case: Testing Local Configurations

Developers can also use stn up to test their local agent configurations in an isolated container environment:

# Test your local workspace in a container
stn up --workspace ~/my-agents

# Test with a specific environment
stn up --environment production

# Test with development tools enabled
stn up --develop

This lets you validate that your agents work correctly in the same containerized environment they'll run in production.

Quick Commands

# Start Station (interactive setup on first run)
stn up

# Start with specific AI provider
stn up --provider openai --model gpt-5

# Check status
stn status

# View logs
stn logs -f

# Stop Station
stn down

# Stop and remove all data (fresh start)
stn down --remove-volume

What stn up Provides

Service	Port	Description
Web UI	8585	Configuration interface
MCP Server	8586	Main MCP endpoint for tools
Agent MCP	8587	Dynamic agent execution
Jaeger UI	16686	Distributed tracing

See Container Lifecycle for detailed architecture.

---

AI Provider Authentication

Station supports multiple authentication methods for AI providers.

API Keys (Default)

The simplest way to authenticate - set your API key as an environment variable:

# CloudShip AI (Recommended - auto-detected when registration key is set)
export CLOUDSHIPAI_REGISTRATION_KEY="csk-..."
# Or: export STN_CLOUDSHIP_KEY="csk-..."

# OpenAI
export OPENAI_API_KEY="sk-..."

# Google Gemini
export GEMINI_API_KEY="..."

# Anthropic (API billing)
export ANTHROPIC_API_KEY="sk-ant-api03-..."

Anthropic OAuth (Claude Max/Pro Subscription) - ⚠️ DEPRECATED

⚠️ DEPRECATED: Anthropic OAuth is currently unavailable.

Anthropic has restricted third-party use of OAuth tokens. This authentication method is not working until further notice.

Use these alternatives instead:

• OPENAI_API_KEY for OpenAI models (recommended)
• GEMINI_API_KEY for Google Gemini models
• ANTHROPIC_API_KEY for Anthropic API (pay-per-token billing)