MCP Server

The @curate-me/mcp-server package exposes Curate-Me platform governance capabilities as Model Context Protocol (MCP)  tools. It runs inside runner containers alongside OpenClaw agents, giving them self-awareness about budgets, fleet peers, session memory, and the ability to communicate with other agents.

Overview

When an OpenClaw agent runs inside a Curate-Me managed runner, the MCP server provides 21 tools across two tiers:

• Tier 1 — Foundation (16 tools): Budget, status, events, skills, peers, A2A, approvals, health, session memory, memory search, desktop screenshots, timeline recording, credentials, extended thinking, and session recordings
• Tier 2 — Collaboration (5 tools): Subagent spawning, agent handoff, fleet messaging, task dispatch, and conversation persistence

OpenClaw Agent → MCP Server (stdio) → Gateway REST API → Control Plane

Installation

The MCP server is pre-installed in all Curate-Me runner container images. For manual installation:

npm install -g @curate-me/mcp-server

Configuration

The MCP server reads configuration from environment variables, automatically injected into runner containers at startup:

OpenClaw Configuration

Add the MCP server to your OpenClaw config’s MCP server list:

{
  "mcpServers": {
    "curate-me": {
      "command": "curate-me-mcp",
      "args": []
    }
  }
}

Tier 1: Foundation Tools (16)

curate_check_budget

Check remaining daily LLM budget for this organization. Returns current spend, daily limit, remaining budget, request count, and whether budget is exceeded.

curate_get_status

Get this runner’s current status including state, resource usage, installed skills, and active session info.

curate_emit_event

Emit an audit event to the control plane. Events are recorded in the immutable audit trail and visible in the dashboard.

curate_install_skill

Install a skill pack on this runner. Skill packs provide additional capabilities (MCP servers, tools, integrations).

curate_list_skills

List all available skill packs in the gallery, or list skills installed on this runner.

curate_discover_peers

Discover other runners in the fleet. Returns online peers with their capabilities and current workload.

curate_send_a2a_task

Send a task to another agent via the A2A (Agent-to-Agent) protocol. The task goes through governance checks (rate limits, PII scanning, cost caps).

curate_request_approval

Request human-in-the-loop (HITL) approval before proceeding with a sensitive action. Creates an approval request in the dashboard queue.

curate_gateway_health

Check the health of the Curate-Me gateway and connected services. No parameters required.

curate_session_memory

Restore context from prior runner sessions. Returns conversation summaries, key findings, and carried-forward data from previous runs. Enables stateful multi-run workflows.

curate_memory_search

Search across all runner session memories for relevant context. Enables knowledge retrieval from past executions across the fleet.

curate_desktop_screenshot

Capture a screenshot of the runner’s desktop for visual debugging. Returns base64-encoded image data.

curate_timeline_record

Record an event into the session timeline for time-travel debugging. Enables replay and inspection of agent execution in the dashboard.

curate_credential_store

Store, list, or delete credentials in the encrypted credential vault.

curate_extended_thinking

Get, set, or check usage of extended thinking configuration.

curate_recording_save

Save or list session recordings (rrweb events) for visual replay.

Tier 2: Collaboration Tools (5)

curate_subagent_spawn

Create, list, or view policy for subagent specs. Each subagent has its own model, tools, cost limit, and conversation budget.

curate_agent_handoff

Hand off work from this agent to another with full context (findings, artifacts, instructions). Enables multi-step workflow pipelines.

curate_fleet_message

Send a message to other agents in the fleet. Supports broadcast to all agents or direct messaging.

curate_task_dispatch

Poll for dispatched tasks from the dashboard or report task completion.

curate_conversation

Store or retrieve conversation entries for audit trail and context continuity.

Usage Examples

Budget-Aware Agent

// Check budget before expensive operation
const budget = await mcp.call("curate_check_budget");
if (budget.budget_exceeded) {
  await mcp.call("curate_emit_event", {
    event_type: "budget_exceeded",
    detail: { action: "skipped_expensive_analysis" }
  });
  return "Budget exceeded - deferring to next cycle";
}

Stateful Multi-Run Agent

// Restore context from prior session
const memories = await mcp.call("curate_session_memory", {
  template_id: "daily_report",
  limit: 1
});
 
// Search for specific knowledge
const insights = await mcp.call("curate_memory_search", {
  query: "customer objections pricing"
});
 
// Record decisions in timeline
await mcp.call("curate_timeline_record", {
  event_type: "decision",
  input_data: { context: "Found 3 prior insights" },
  output_data: { decision: "Using revised pricing strategy" }
});

Fleet Collaboration

// Discover peers
const peers = await mcp.call("curate_discover_peers", { state_filter: "online" });
 
// Delegate sub-task
await mcp.call("curate_send_a2a_task", {
  to_agent: peers[0].runner_id,
  content: "Analyze the uploaded CSV and generate a summary report",
  priority: "normal"
});
 
// Broadcast learning to fleet
await mcp.call("curate_fleet_message", {
  content: "Found that CSV headers must be normalized before processing",
  broadcast: true
});

Multi-Step Workflow Pipeline

// Agent A completes analysis, hands off to Agent B
await mcp.call("curate_agent_handoff", {
  to_agent: "runner_compliance_reviewer",
  summary: "Completed financial analysis for Q4",
  findings: ["Revenue up 15%", "Costs down 8%", "New market opportunity in APAC"],
  artifacts: ["/workspace/q4_analysis.pdf", "/workspace/charts.xlsx"],
  instructions: "Review for regulatory compliance before publishing"
});

Secure Credential Management

// Store API key securely
await mcp.call("curate_credential_store", {
  action: "store",
  name: "stripe_api_key",
  credential_type: "api_key",
  provider_hint: "stripe",
  value: "sk_live_xxx"
});
 
// List stored credentials
const creds = await mcp.call("curate_credential_store", { action: "list" });

HITL for Sensitive Operations

// Request approval before destructive action
await mcp.call("curate_request_approval", {
  action_type: "delete_data",
  description: "Delete 1,247 stale user records older than 2 years",
  estimated_cost_usd: 0
});

Architecture

┌─────────────────────────────────────────┐
│           Runner Container              │
│                                         │
│  ┌──────────┐     ┌──────────────────┐  │
│  │ OpenClaw │────→│ MCP Server       │  │
│  │ Agent    │stdio│ (21 tools)       │  │
│  └──────────┘     └────────┬─────────┘  │
│                            │ HTTP       │
└────────────────────────────┼────────────┘
                             │
                    ┌────────▼────────┐
                    │   Gateway :8002 │
                    │ (governance     │
                    │  chain + proxy) │
                    └────────┬────────┘
                             │
                    ┌────────▼────────┐
                    │  Control Plane  │
                    │  (MongoDB +     │
                    │   Redis)        │
                    └─────────────────┘

Tool Categories

Development

cd packages/mcp-server
 
# Install dependencies
npm install
 
# Run the server (for testing)
npm start
 
# Development with auto-reload
npm run dev

Testing MCP Tools Locally

# Start the gateway first
cd services/backend
poetry run uvicorn src.main_gateway:app --reload --port 8002
 
# Test the MCP server — list all tools
echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"1.0.0"}}}
{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}' | node packages/mcp-server/src/index.js
 
# Call a specific tool
echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"1.0.0"}}}
{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"curate_gateway_health","arguments":{}}}' | CM_GATEWAY_URL=http://localhost:8002 node packages/mcp-server/src/index.js

Related

• CLI Reference - Command-line tool also installed in containers
• Managed Runners - Runner container lifecycle and management
• AI Gateway - The governance proxy that MCP tools communicate with