Introducing Cencori MCP

The tools an agent can reach define what it can become.

Today we are launching Cencori MCP, the official Model Context Protocol server for the Cencori platform.

It gives Cursor, Claude Desktop, Codex, and any MCP-compatible client a direct, structured way to search Cencori Web, read our documentation, inspect gateway health and usage, work with memory and agents, manage durable sessions, query governance, and run multimodal inference.

One server. One project key. The Cencori platform, available as tools.

npx -y @cencori/mcp@latest

Cencori MCP is useful before you add a key.

Run the package and your agent can immediately search the Cencori documentation, retrieve an exact page, inspect the docs navigation, or load our complete llm.txt integration contract.

{
  "mcpServers": {
    "cencori": {
      "command": "npx",
      "args": ["-y", "@cencori/mcp@latest"]
    }
  }
}

That gives the client four public tools:

• search_docs
• get_doc
• list_docs
• get_integration_guide

It also exposes guidance tools for the decisions that must remain human-controlled: creating or revoking API keys, changing billing, managing access, and activating governance policy.

An agent can tell you exactly what to do. It cannot quietly do those things for you.

Add a Cencori project key and the server exposes our first-party Web layer:

{
  "mcpServers": {
    "cencori": {
      "command": "npx",
      "args": ["-y", "@cencori/mcp@latest"],
      "env": {
        "CENCORI_API_KEY": "csk_your_project_key"
      }
    }
  }
}

Now the client can:

• web_search across Cencori's owned hybrid index
• web_fetch a bounded public text resource
• web_extract clean content, links, metadata, dates, and evidence spans
• get_web_browser_job to poll a JavaScript browser exploration

Cencori Web does not forward these queries to a hosted search provider. The crawler, corpus, local embeddings, retrieval system, ranking pipeline, extraction layer, and browser workers belong to Cencori.

Search results return more than a title and a URL. They include an evidence quote, content hash, and retrieval timestamp so the agent can show what supported its answer and when that evidence was observed.

Every Web tool is marked as open-world access. Every description tells the client that retrieved content is untrusted data, never instructions. A page can inform the agent; it cannot promote itself into the agent's control plane.

With the same project key, an agent can inspect the infrastructure its work runs through.

Gateway

• List available models
• Inspect request, token, cost, and latency metrics
• Check platform health
• Check project quota

Memory

• List and semantically search memories
• Read a specific memory
• Inspect resolved entities and the memory graph
• Review forget suggestions

Agents and sessions

• List and inspect agents
• Poll agent actions
• Read durable sessions and their event streams

Governance

• Inspect policies and roles
• Review change requests
• Read the governance ledger and evidence
• Discover available policy templates

This is not a second implementation of Cencori inside an MCP package. The server is a thin adapter over the same public APIs applications use. Authentication, tenant isolation, quotas, governance, and audit logging remain enforced by the platform.

Agent tools should not all have the same power.

Cencori MCP divides operations into explicit action tiers:

Without a write flag, the server does not register tools that incur inference cost, queue browser work, crawl into a project index, or change project state.

To enable the full non-destructive surface:

{
  "mcpServers": {
    "cencori": {
      "command": "npx",
      "args": ["-y", "@cencori/mcp@latest"],
      "env": {
        "CENCORI_API_KEY": "csk_your_project_key",
        "CENCORI_MCP_WRITE": "1"
      }
    }
  }
}

That adds:

• JavaScript browser exploration and project-scoped crawling
• Text generation and RAG
• Embeddings and moderation
• Image generation and vision
• Document extraction, summarization, and Q&A
• Text-to-speech and transcription
• Memory, agent, session, and governance-draft writes

Destructive tools remain absent until you enable them separately.

MCP clients receive readOnlyHint, destructiveHint, and openWorldHint metadata on each tool. The client can understand the nature of an operation before it asks to execute it.

You can narrow the tool surface with CENCORI_MCP_FEATURES:

{
  "CENCORI_MCP_FEATURES": "web,docs"
}

Available feature groups are:

docs, guidance, gateway, agents, memory,
sessions, web, governance, multimodal

An agent researching technical documentation does not need billing metrics. An operations agent may need gateway and governance reads but no browser. A product agent may need memory and sessions but not destructive access.

The smallest sufficient tool surface is usually the best one.

APIs are designed for applications that already know which endpoint to call. MCP lets an agent discover the available operations, understand their schemas, and select the right tool during a task.

That distinction matters. A developer can now ask:

Search Cencori Web for the latest PostgreSQL row-level security guidance.
Compare the strongest three sources, quote the evidence you used, and include
each source URL and retrieval timestamp.

Or:

Check our Cencori gateway health and seven-day latency. Then inspect the active
governance policies and tell me whether anything would block a document-analysis
agent from entering production.

The agent sees typed tools. Cencori sees authenticated, scoped API requests. The user sees an inspectable result.

Cencori MCP is open source and available on npm as @cencori/mcp@0.7.1.

npx -y @cencori/mcp@latest

Start without a key for documentation. Add CENCORI_API_KEY for Web and platform reads. Enable write or destructive capabilities only when the agent's job requires them.

Agents should be able to understand and operate the infrastructure they run on. Now Cencori speaks their native protocol.

Read the MCP documentation → · Explore Cencori Web → · View the package on npm →