ThinkNEO MCP servers: 68 tools you can use today

Nine ThinkNEO MCP servers (main control plane + 8 SMB MCPs) offering 68 tools across guardrails, trust scoring, observability, routing, A2A, compliance, cost copilot, and monitor. All auth via EMA, all audited, all scoped by risk tier.

This post inventories the MCP servers ThinkNEO operates in production, their tools, their trust tier, and how to consume them from your AI application. There are nine servers, 68 tools total, all governed by the same enforcement plane that fronts the ThinkNEO LLM gateway.

Discovery

All ThinkNEO MCP servers are listed with authoritative ownership on Glama, with the main control-plane server holding AAA verification. Direct HTTPS URLs are stable; the registry entry is for discovery, not the connection.

The nine servers

1. Main control-plane MCP

URL: https://mcp.thinkneo.ai

Scope: tenant-facing operations — list projects, mint keys, read metering, adjust budgets, manage models. This is the server your dashboard AI or IDE-integrated agent talks to for administrative workflows.

Auth: EMA-compliant OAuth (via your IdP), or ThinkNEO admin token (Bearer).

2. Guardrails MCP (SMB)

URL: https://mcp-guardrails.thinkneo.ai

Scope: content policy inspection — PII detection, jailbreak markers, credential pattern matching, toxicity thresholds. Callable both from the ThinkNEO gateway and directly from your applications.

3. Trust Score MCP (SMB)

Scope: computes and returns a trust score for a call context, based on identity, historical behavior, and content signals. Used by the gateway for high-risk decisions; also callable by your applications for pre-flight risk checks.

4. Observability MCP (SMB)

Scope: query the audit ledger and metering tables — "show me the last 100 denied calls in workspace X," "how much did project Y spend this month broken down by model." Read-only by design; enforcement decisions belong to the gateway.

5. Router MCP (SMB)

Scope: given a task description and constraints (max cost, max latency, quality target), returns a routing recommendation (which model, which provider, which BYOK key). Advisory — the caller decides whether to accept.

6. A2A-Lite MCP (SMB)

Scope: agent-to-agent orchestration primitives — task handoff, capability discovery, session context sharing. Interoperates with the Google A2A protocol.

7. Compliance MCP (SMB)

Scope: on-demand compliance evaluation of a governed project — SOC 2 control coverage, ISO 42001 mapping, audit-ledger integrity report. Observes and opines; never decides.

8. Cost Copilot MCP (SMB)

Scope: cost analysis and savings recommendations grounded in arithmetic (real tokens × real prices). LLM-assisted description of findings; savings numbers themselves are deterministic.

9. Monitor MCP (SMB)

Scope: rule-based alert evaluation — reads the audit ledger and fires configured alert channels when thresholds are crossed. Deterministic detector is truth; LLM-assisted phrasing of recommendations is disclosed.

How to consume them

From any MCP-capable client (Claude Desktop, Cursor, VS Code with MCP extension, your own agent framework):

{
  "mcpServers": {
    "thinkneo": {
      "url": "https://mcp.thinkneo.ai",
      "auth": {"type": "oauth", "provider": "your-idp"}
    },
    "thinkneo-guardrails": {
      "url": "https://mcp-guardrails.thinkneo.ai",
      "auth": {"type": "oauth", "provider": "your-idp"}
    }
  }
}

Add the servers you need. Do not add all nine by default — least privilege applies here too.

What the governance guarantees

Every tool call is authenticated by the same OAuth layer that fronts the LLM plane.
Every tool call produces an audit row in the same immutable ledger.
Read tools (observability, compliance report, cost analysis) require the mcp.read scope.
Write tools (mint key, adjust budget, enable model) require the mcp.write scope plus the caller's role permitting that action in the target workspace.
Destructive tools (revoke key, delete project) require explicit approval flow per invocation.
Every server enforces its own rate limit (100 rps read, 10 rps write, 1 rps destructive).
Credentials never appear in MCP payloads. If your prompt to an LLM contains tnk_... or Bearer ..., the guardrail drops the context before egress.

Roadmap

2026-Q3: Endpoint-Control MCP (device-scoped policy issuance), currently dormant in production.

2026-Q4: BYOK key rotation MCP (one-command rotate-new-first across all registered providers).

These will follow the same auth, scope, and audit discipline.

How this compares to competitors

Most MCP catalogs advertise a large count of tools without a governance framing. ThinkNEO's 68 tools are inventoried, scoped, rate-limited, and auditable from day one. That is the difference between a marketplace and a governed integration surface. Both models have their place; enterprise consumers usually need the latter.