MCP MCP Server

@replenishradar/mcp-server

Connect AI agents to live ReplenishRadar inventory data. Works with Claude Desktop, OpenClaw, and any MCP-compatible client.

Setup

1. Get your API key from ReplenishRadar > Settings > API Keys.
2. Add this server to your MCP client config. In Claude Desktop, open Settings > Developer > Edit Config:

{
  "mcpServers": {
    "replenishradar": {
      "command": "npx",
      "args": ["-y", "@replenishradar/mcp-server"],
      "env": {
        "REPLENISHRADAR_API_KEY": "rr_sk_your_key_here"
      }
    }
  }
}

1. Restart Claude Desktop and ask: "What are my top stockout risks?"

No ReplenishRadar account yet? Use the public agent intake endpoint to request a short-lived, read-only sandbox key before configuring the server:

POST https://api.replenishradar.com/api/public/agent-intake

Available Tools

Read

Standard includes limited read API access. Growth adds diagnosis/status read tools.

Basic write (Standard tier and up)

Safe, low-blast-radius writes - no money or stock mutation.

Sensitive write (Growth tier and up)

Money / stock mutation.

PO creation always starts as a draft. rr_send_purchase_order only sends a PO that was already human-approved.

Action execution has two modes. rr_execute_replenishment_action with mode="draft" (the default) creates a draft PO and/or draft transfer for an action's items and never auto-sends to a supplier or commits stock. With mode="autonomous" it additionally sends the PO or commits the transfer, but only when the key has sensitive_write + the operation's tool group + an enabled per-key budget (autonomous_enabled=true) and every guardrail passes; otherwise it leaves a draft and returns an rr_request_approval hint (draft_fallback) or refuses without mutating (hard_refuse). Call rr_prepare_replenishment_action first to get the action's current updated_at, then pass it as expected_updated_at so a stale action is rejected before any write. Pass a stable idempotency_key so a retried call replays the original result instead of double-creating artifacts or double-sending.

Agent context store

A small, org-scoped, auditable key/value memory so an agent can remember bounded workflow state. rr_get_agent_context / rr_list_agent_context are read-capability; rr_set_agent_context / rr_delete_agent_context are basic-write. All four are in the agent_context tool group.

Rules and limits:

• Namespaces are code-owned and per-key allowlisted. A human admin must grant your API key read/write/delete on each namespace in Settings > API Keys. Same-organization access alone is not sufficient.
• No secrets, no raw PII. Writes that look like tokens, passwords, private keys, credentials, or raw emails/phones/addresses are rejected.
• value must be a JSON object (not a scalar or array), capped at 16 KiB serialized.
• ttl_seconds is optional and capped at 90 days; with no TTL the entry persists while the organization exists.
• rr_list_agent_context requires namespace and never lists across namespaces.
• scope_ref is descriptive metadata only and is not part of uniqueness - encode per-scope entries into the key (for example supplier_rules/sku:ABC123).
• Sensitive namespaces (for example sourcing_economics) additionally require the key to hold the sensitive economics read capability + tool group.

Spend budgets (admin-managed, no MCP tool)

Per-API-key spend budgets - per-PO value cap, rolling-24h daily cap, max transfer quantity delta, an autonomous_enabled flag, and vendor / destination / transfer-source (or route-pair) allowlists - constrain what an autonomous key may buy or move. They are managed only by a human admin in Settings > API Keys (a JWT-admin Settings API); there is no MCP tool to read or change a budget, and an API key cannot reach the budget routes, not even for its own caps. Budgets are consumed only by autonomous send/commit, autonomous mode is off by default, and a denied autonomous action receives only a short decision reason - never the cap amounts, the allowlists, or another key's ledger.

Rate Limits

• Standard ($99/mo): 10 calls/hour, read + basic-write tools
• Growth ($199/mo): 100 calls/hour, read + basic + sensitive read/write + diagnosis/status tools
• Scale ($499/mo): 1,000 calls/hour, full read + write capability set

Learn More

• MCP Setup Guide
• ReplenishRadar