@savantoai/mcp-server

A local Model Context Protocol server that exposes your Savanto AI workspace to Claude, ChatGPT, Cursor, and any other MCP-compatible client — so you can configure, populate, and operate your store's AI assistant by talking to your own AI, instead of clicking through a dashboard.

What it does

Once configured, your agent gains a curated set of tools that mirror the Savanto REST API, spanning the full configure → observe → refine loop:

| Category | Representative tools | Scope | | --- | --- | --- | | Workspaces | list_workspaces, create_workspace, update_workspace, delete_workspace | tenant:admin | | Configuration | get_workspace_settings, update_workspace_settings, custom-domain CRUD, discover_tools, generate_domain_config, validate_custom_domain, test_domain_connection, generate_color_scheme, chat/search widget config | config:admin | | Content | upsert_product/upsert_post (+ bulk_, list_, get_, patch_, delete_*) | admin:products, admin:posts | | Taxonomies | upsert_taxonomy, bulk_upsert_taxonomies, list/get/delete_taxonomy | admin:taxonomies | | Prompts | upsert_prompt, list_prompts, search_prompts, delete_prompt (+ bulk) | admin:prompts, prompts:read | | Webhooks | create_webhook, list/get/update/delete_webhook, test_webhook, get_webhook_stats | admin:webhooks | | Crawl | start_crawl, get_crawl_status/history/config, update_crawl_config | admin:posts | | Search | search_products, search_posts | search:products, search:posts | | Analytics | get_search_analytics, get_chat_analytics, get_feedback_analytics, search_search_logs, list_feedback | tenant:admin, feedback:admin | | Threads | search_threads, get_thread, get_thread_messages, get_thread_analytics, delete_thread, bulk_delete_threads | threads:admin | | Chat | chat | chat | | Diagnostics | whoami, get_tenant_usage | (none) / tenant:admin |

Two things keep the surface safe and legible to clients:

Scope-gated at startup — the server probes /tenant/whoami and only registers tools your key can actually use. An agent is never shown a tool it would get a 403 for, and a publishable widget key sees almost nothing.
Annotated — every tool carries MCP hints (readOnlyHint, destructiveHint, idempotentHint) so clients can auto-approve safe reads and flag destructive writes; deletes additionally require an explicit confirm: true.

The server also exposes Skills (MCP prompts) — step-by-step playbooks for common multi-tool workflows:

onboard-store-end-to-end – create a workspace, ingest content, configure behaviour + branding, smoke-test
onboard-wordpress / onboard-shopify – platform-specific onboarding walkthroughs
configure-chat – tune persona, special instructions, and handoff rules
configure-custom-domain – wire a custom capability (order tracking, account lookup) to MCP servers / REST APIs
audit-and-improve – the observe→refine loop: find failing chats / zero-result searches / negative feedback and fix them
debug-empty-search – diagnose why a product search returns no hits
migrate-from-competitor – bulk-import from another chat vendor's export

Requirements

Node.js 20 or later
A secret Savanto API key (starts with if_sk_…). Create one in the API Keys page of your dashboard.

Publishable keys (if_pk_…) are client-side and cannot provision workspaces — the server will refuse to start with one.

Quick start

No global install needed — run it with npx:

export SAVANTO_API_KEY=if_sk_your_key_here
npx -y @savantoai/mcp-server

Point to a non-production cloud (staging, local dev):

export SAVANTO_API_URL=http://localhost:3001

Remote server (preview)

In addition to the local stdio server above, the same tool surface can run as a hosted HTTP server so clients connect to a URL instead of spawning npx — no local Node, no per-machine config. This is the path toward one-click "Connect to Claude/ChatGPT" (OAuth) onboarding; today it accepts your secret key as a Bearer token.

# Each client authenticates per-request — there is NO server-wide key.
SAVANTO_API_URL=https://api.savanto.ai PORT=8080 npx -y -p @savantoai/mcp-server savanto-mcp-http

The server mounts the MCP endpoint at /mcp and a liveness probe at /healthz. Clients send their key as Authorization: Bearer if_sk_…; the tool surface is scope-gated to that key's tenant, exactly as in the stdio server. Point an MCP client that supports remote (Streamable HTTP) servers at https://your-host/mcp with that bearer token.

Auth is currently the raw secret key. A future release replaces it with OAuth-issued, tenant-scoped tokens so customers can connect with zero key handling — the transport and tool layer are unchanged by that swap.

Client configuration

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "savanto": {
      "command": "npx",
      "args": ["-y", "@savantoai/mcp-server"],
      "env": {
        "SAVANTO_API_KEY": "if_sk_your_key_here"
      }
    }
  }
}

Restart Claude Desktop. You should see a hammer/tool icon in the message bar; the Savanto tools are listed there.

Cursor

In Cursor settings → Features → Model Context Protocol → Add new MCP server:

{
  "savanto": {
    "command": "npx",
    "args": ["-y", "@savantoai/mcp-server"],
    "env": { "SAVANTO_API_KEY": "if_sk_your_key_here" }
  }
}

Cline / Roo / other VS Code agents

Add to the extension's MCP config (usually a JSON file under ~/.cline or similar):

{
  "mcpServers": {
    "savanto": {
      "command": "npx",
      "args": ["-y", "@savantoai/mcp-server"],
      "env": { "SAVANTO_API_KEY": "if_sk_your_key_here" }
    }
  }
}

OpenAI Agents / Responses API

from openai import OpenAI
from mcp import StdioServerParameters

server = StdioServerParameters(
    command="npx",
    args=["-y", "@savantoai/mcp-server"],
    env={"SAVANTO_API_KEY": "if_sk_your_key_here"},
)

Local MCP Inspector (for debugging)

npx @modelcontextprotocol/inspector npx @savantoai/mcp-server

The Inspector gives you a web UI to list tools, call them directly, and watch request/response payloads — great for confirming your key is wired correctly before handing the server to an agent.

Example prompts

Once the server is registered in your MCP client, try:

"Set up a new Savanto workspace for acme-store, crawl https://acme.test, give it an outdoor-adventure tone, and brand the widget around #0a7d2c." (end-to-end onboarding)

"Look at acme-store's last 30 days — what are visitors searching for that returns nothing, and which conversations went unresolved? Then add content to fix the top few." (the observe→refine loop)

"Add an order-tracking capability to acme-store backed by our MCP server at https://mcp.acme.test/orders, validate it, and test it before enabling." (custom domain)

"Why did this conversation get a thumbs-down?" — pull list_feedback, read the thread, and propose a fix.

The agent picks the right tools automatically (and clients can auto-approve the read-only ones). You can also invoke a Skill explicitly — e.g. in Claude Desktop, /onboard-store-end-to-end or /audit-and-improve kicks off that full playbook.

Environment variables

| Variable | Default | Purpose | | ----------------- | ------------------------------- | ------------------------------------------------ | | SAVANTO_API_KEY | _(required)_ | Your secret API key (if_sk_…). | | SAVANTO_API_URL | https://api.savanto.ai | Override for staging / local dev. |

Security

Always use separate API keys per agent / machine — so you can revoke one without affecting the others. The API Keys page tracks the last-used timestamp of each key.
Keys are passed via environment variables, never logged. The server prints a one-line identity banner on startup (to stderr) showing the tenant id and scope list — no secrets.
The server runs over stdio and never opens a network port. It only speaks to the Savanto API host you point it at.
delete_workspace requires an explicit confirm: true parameter in the tool call — a safety gate against hallucinated destructive operations.

Local development

From the repo root:

npm install
npm run build --workspace=@savantoai/mcp-server
SAVANTO_API_KEY=if_sk_… SAVANTO_API_URL=http://localhost:3001 node sdks/mcp/dist/stdio.js

Run the tests:

npm run test --workspace=@savantoai/mcp-server

License

MIT. See LICENSE.

mcp-server