@relayplane/mcp-server

MCP server for efficient AI workflow orchestration with RelayPlane.

Why RelayPlane?

90%+ context reduction on multi-step pipelines.

When AI agents use MCP tools directly, they face two problems:

Tool definitions bloat context — 100k+ tokens for many tools
Intermediate results pass through context — a 50k token document copied between steps

RelayPlane workflows solve this by keeping intermediate results in the workflow engine, not your context window.

Installation

# Add to Claude Code
claude mcp add relayplane -- npx @relayplane/mcp-server

# With custom output directory for generated code
claude mcp add relayplane -- npx @relayplane/mcp-server --out-dir ./src/lib/relayplane

# With budget override
claude mcp add relayplane -- npx @relayplane/mcp-server --max-daily-cost 10

Environment Variables

Set your provider API keys:

export OPENAI_API_KEY=sk-...
export ANTHROPIC_API_KEY=sk-ant-...
export GOOGLE_API_KEY=...
export XAI_API_KEY=...

Available Tools

Tool	Purpose	Cost
`relay_run`	Test single prompt	Provider cost
`relay_workflow_run`	Test multi-step workflow	Provider cost
`relay_models_list`	List available models	Free
`relay_workflow_validate`	Validate DAG structure (no LLM calls)	Free
`relay_skills_list`	Discover pre-built skills	Free
`relay_runs_list`	List recent runs	Free
`relay_run_get`	Get run details + trace URL	Free

Usage Example

// Single model call
const result = await relay_run({
  model: "openai:gpt-4o",
  prompt: "Summarize this document...",
  systemPrompt: "You are a helpful assistant."
});

// Multi-step workflow
const workflow = await relay_workflow_run({
  name: "extract-and-validate",
  steps: [
    {
      name: "extract",
      model: "openai:gpt-4o",
      prompt: "Extract all invoice fields from: {{input.document}}"
    },
    {
      name: "validate",
      model: "anthropic:claude-3-5-sonnet-20241022",
      depends: ["extract"],
      prompt: "Validate the extracted data: {{steps.extract.output}}"
    }
  ],
  input: {
    document: "Invoice #123..."
  }
});

console.log(workflow.contextReduction); // "94% (saved ~12k tokens)"

Budget Limits

Budget tracking protects you from runaway provider costs (OpenAI, Anthropic bills).RelayPlane is BYOK — we don't charge for API usage.

Default limits:

Daily: $5.00
Single call: $0.50
Calls per hour: 100

Override with CLI flags:

--max-daily-cost 10
--max-single-call-cost 1
--max-calls-per-hour 200

Pre-built Skills

Discover pre-built workflow patterns:

relay_skills_list({ category: "extraction" })

Available skills:

invoice-processor - Extract, validate, summarize invoices (97% context reduction)
content-pipeline - Research, draft, refine content (85% context reduction)
lead-enrichment - Enrich and score leads (90% context reduction)
document-qa - RAG-based document Q&A (75% context reduction)
code-review - Multi-stage code review (80% context reduction)
email-classifier - Classify and route emails (95% context reduction)

Configuration

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

{
  "codegenOutDir": "./servers/relayplane",
  "maxDailyCostUsd": 5.00,
  "maxSingleCallCostUsd": 0.50,
  "maxCallsPerHour": 100,
  "providers": {
    "openai": { "apiKey": "${OPENAI_API_KEY}" },
    "anthropic": { "apiKey": "${ANTHROPIC_API_KEY}" }
  }
}

Development

# Build
pnpm build

# Run locally
pnpm dev

# Test
pnpm test

License

MIT

@relayplane/mcp-server