For teams tired of duct-taping "yet another database plugin"

MCP Connectors

Контракт

When an agent needs shipment status, it invokes get_order_status instead of generating raw SQL. Permissions apply strictly to tables and columns you explicitly allow. Secrets stay in the vault[1].

Запрещено
SELECT * FROM orders
Инструмент
get_order_status({ tracking: "84920" })
Поле
in_transit · ETA Thu 16:40 · city: hub-west

Scope

Read-only access strictly to selected fields

Expose specific views and columns, never entire schemas. The agent cannot see card PANs even if present in the database table.

A connector's scope must never exceed an employee's access tier.

PII

PII Masking

  • PII is stripped before entering the prompt

    Data masking happens pre-inference. GDPR compliance requires strict data minimization and controlled access boundaries.

  • Write operations are isolated actions with human approval

    MCP exposes the tool. Invocation policies, idempotency keys, and Human-in-the-Loop (HITL) safeguards govern agentic actions.

FAQ

Frequently Asked Questions Before Granting Database Access

What is MCP in volbor?

MCP (Model Context Protocol) is an open standard allowing AI agents to access vetted external tools over a structured schema. It is neither arbitrary SQL access nor a replacement for Core APIs.

Why use MCP instead of conventional function calling?

MCP is an open standard. A single tool server works across different clients and LLM providers, avoiding proprietary vendor lock-in.

Can I connect an internal on-premise service?

Yes. Using official SDKs (TypeScript, Python), you can run your own MCP server and connect it to your workspace. You retain full control over scopes and permissions.

Where are passwords and connection strings stored?

Inside your workspace secrets vault. The LLM only sees tool signatures and parameter schemas—never DSNs or database credentials.

How do we prevent overloading our production database?

Through rate limiting, connection pooling, read caching, and barring arbitrary SQL queries. Heavy reporting queries should always route to read replicas, not OLTP masters.

What happens if the database goes down?

The agent receives a structured error, informs the customer transparently, and escalates to a human rep. It never hallucinates an answer.

Do I need an Anthropic commercial license to use MCP?

No. MCP is an open-source specification published by Anthropic; using it within volbor does not require an Anthropic license.

Does the agent write directly to database tables?

No. Connectors are read-only by default. Write operations require explicit agentic actions backed by Human-in-the-Loop (HITL) approval.

Let shipment statuses come from live fields, not canned guesswork

One read-only connector, three scoped columns, interactive sandbox. Pasting SQL into prompts is no longer acceptable.

Methodology & Sources

How Metrics on This Page Are Calculated

  1. [1]
    MCP is an open protocol for AI tool access

    The Model Context Protocol is an open standard (Anthropic, Nov 2024; supported by an open ecosystem of servers and clients). Cited here as an architectural standard separating tools, secrets, and models, not a volbor latency measurement.

  2. [2]
    Catalog connector takes minutes vs. days or weeks for custom wrappers

    Catalog connector: an OAuth token or API secret plus view selection. A custom "bot-to-ERP" microservice requires contract definitions, auth, retries, deployment, and monitoring—typically taking days or weeks of engineering time. This illustrates connection complexity, not an empirical survey.

  3. [3]
    Maintaining a zoo of custom connectors consumes a fraction of an FTE

    Worked example: maintaining three bespoke integrations (wiki, inventory, CRM) occupies a fraction of an engineer's workload. At a loaded cost of $4,000–$7,000/month for an engineer, spending a third of their time on glue code equals $1,300–$2,300/month. Standard connectors remove this glue. This is an operational cost illustration, not volbor pricing or guaranteed savings.