RelayPlane MCP Server

Reduce AI context usage by 90%+ in multi-step workflows

RelayPlane keeps intermediate results in the workflow engine instead of passing them through your context window—saving tokens and reducing costs.

Table of Contents

• Quick Start

• Installation Options

• Model IDs

• Available Tools

• Budget Protection

• Pre-built Skills

• Configuration

• Troubleshooting

Quick Start

1. Install with API Keys (Recommended)

2. Restart Claude Code

Important: You must fully restart Claude Code after adding the MCP server. The /mcp command only reconnects—it doesn't reload environment variables.

3. Test the Connection

Ask Claude: "Use relay_models_list to show configured providers"

Models should show configured: true for providers with valid API keys.

Installation Options

Option A: Inline API Keys (Simplest)

Option B: Shell Environment Variables

First, add to your shell profile (~/.zshrc or ~/.bashrc):

Then source and install:

Option C: Manual Configuration

Edit ~/.claude.json directly:

Warning: The env field must contain actual API keys, not variable references like ${OPENAI_API_KEY}. Variable substitution is not supported in the MCP config file.

Model IDs

Important: Always check https://relayplane.com/docs/providers for the latest model IDs. The relay_models_list tool may return outdated information.

OpenAI — prefix: openai:

Anthropic — prefix: anthropic:

Google — prefix: google:

xAI — prefix: xai:

Example Usage

Available Tools

Budget Protection

Default safeguards (customizable via CLI flags):

RelayPlane is BYOK (Bring Your Own Keys)—we don't charge for API usage. Costs reflect only your provider bills.

Pre-built Skills

Use relay_skills_list to see available workflow templates:

Configuration

Persistent Config File

Create ~/.relayplane/mcp-config.json:

Note: API keys should be passed via environment variables or the Claude Code MCP env field—not stored in this config file.

Troubleshooting

"Provider not configured" Error

Causes:

1. API key not passed to MCP server

2. Claude Code not restarted after config change

Solutions:

1. Check your MCP config in ~/.claude.json:

2. Fully restart Claude Code (exit with Ctrl+C, relaunch)

3. Verify configuration: Ask Claude: "Use relay_models_list and check which show configured: true"

Model Not Found (404 Error)

Cause: Model ID is outdated or incorrect.

Solution: Check current model IDs at: https://relayplane.com/docs/providers

Common fixes:

• Use claude-3-5-haiku-20241022 instead of claude-3-5-sonnet-20241022

• Use gpt-4.1 instead of gpt-4o for latest OpenAI

Config Changes Not Taking Effect

Cause: /mcp reconnect doesn't reload environment variables.

Solution: Fully restart Claude Code:

1. Exit with Ctrl+C

2. Relaunch claude

3. Run /mcp to verify connection

Workflow Validation Passes But Execution Fails

Cause: relay_workflow_validate only checks DAG structure, not:

• API key validity

• Model availability

• Schema compatibility

Solution: Test with a simple relay_run first:

Quick Test

After setup, verify everything works:

Expected: Both steps complete successfully with structured output.

Support

• Documentation: https://relayplane.com/docs

• Model IDs: https://relayplane.com/docs/providers

• Issues: https://github.com/RelayPlane/mcp-server/issues

License

MIT