Agent Context

Agent Context lets AI agents read and query Sanity content via MCP. It's schema-aware: agents can query fields, follow references, and use semantic search.

• Install the skill:

npx skills add sanity-io/agent-context --all

pnpm dlx skills add sanity-io/agent-context --all

yarn dlx skills add sanity-io/agent-context --all

bunx skills add sanity-io/agent-context --all

• Prompt your AI assistant to use the skill based on what you want to build:
  • Adding an agent to an existing application
  • Building a new application with agent capabilities
  • Setting up Agent Context for a specific use case
• Follow the guided setup to achieve your desired configuration.

For more detail, see Using the skill. For manual setup without the skill, see Manual setup.

The Agent Context MCP is a hosted MCP server that lets AI agents read and query a Sanity dataset in a structured, schema-aware way.

It provides:

• Tools that let an agent understand your schema and query content
• Optional content filtering to limit what the agent can access
• Optional semantic search (when embeddings are enabled)

Agent Context is configured by an Agent Context document in your dataset, which defines what the MCP can access. To learn how to configure your Studio to create Agent Context documents, see Studio plugin below.

The Agent Context document has the following fields:

• name: human readable name
• slug: identifier used when connecting to the MCP
• instructions: custom instructions for how AI agents should work with your content E.g., "Always respond in Spanish" or "Focus on product documentation only.
• groqFilter: optional filter that limits what content the agent can access. This should be a GROQ filter (e.g., _type == "product"), not a full GROQ query (e.g., *[_type == "product"]).

Agent Context exposes three tools. These are discovered automatically when you connect an MCP client.

The Agent Context MCP server is hosted at https://api.sanity.io/:apiVersion/agent-context/:projectId/:dataset/:slug, where:

• apiVersion: API version in the format vYYYY-MM-DD (e.g., v2024-01-01)
• projectId: the ID of the project
• dataset: the name of the dataset
• slug: the slug of an Agent Context document which configures the MCP

The MCP URL also accepts query parameters to customize the request. These are runtime options and are not stored in the Agent Context document:

The agentContextPlugin from the @sanity/agent-context package is a Studio plugin that helps you configure Agent Context. It registers the Agent Context document type and adds a custom input for selecting which document types the agent can access.

Three skills are available to help you build and refine agents using Agent Context:

create-agent-with-sanity-context The primary skill. It walks you through connecting an existing agent to Agent Context MCP, building a new project from scratch with a frontend and Studio, or adding an agent to an existing project. This includes installing and configuring the @sanity/agent-context Studio plugin, creating an Agent Context document, generating a reference agent implementation, and adding a basic chat UI.

shape-your-agent An optional interactive session that helps you craft a system prompt for your agent. It guides you through defining your agent's audience, tone, boundaries, and fallback behavior, then produces a focused prompt (typically 200-400 words) tuned for use with Agent Context MCP.

dial-your-context An interactive session that helps you write the Instructions field for an Agent Context document. The Instructions field provides information the agent can't determine from the schema alone, such as misleading field names, required filters, non-obvious relationships between types, and useful query patterns.

Before you begin, ensure you have:

• A Sanity project with content you want the agent to access
• Sanity Studio v5.1.0+ or later - required for server-side schema, which Agent Context depends on
• Schema deployment - Opening the Studio in a browser triggers schema deployment. The Studio must be deployed, not running on localhost.
• A Sanity API read token - create one at sanity.io/manage > Project > API > Tokens
• An LLM API key - from Anthropic, OpenAI, or another provider
• A frontend application (optional) - where your agent will live (Next.js, Remix, etc.). The skill can help set this up if needed.

There are two paths to get started:

• Using the skill - guided setup that helps you build an agent from scratch
• Manual setup - do the setup yourself without the skill

Best for building a new agent from scratch. The skill guides you through Studio setup, agent implementation, and frontend UI.

• Install the skill:

npx skills add sanity-io/agent-context --all

pnpm dlx skills add sanity-io/agent-context --all

yarn dlx skills add sanity-io/agent-context --all

bunx skills add sanity-io/agent-context --all

• Prompt your AI assistant:

Use the create-agent-with-sanity-context skill to help me build an agent in this project.

The skill guides you through:

• Studio setup: install and configure @sanity/agent-context, then create Agent Context documents and get an MCP URL.
• Agent implementation: generate a reference agent implementation (for example Next.js + Vercel AI SDK) that you can adapt to other stacks.
• Frontend UI: if applicable, add a basic UI for interacting with the agent (for example a chat interface).

Full setup if you're building from scratch without the skill.

• Ensure you're running Sanity Studio v5.1.0+ or later. Agent Context requires server-side schema, which is only available in v5.1.0+.
• Install the @sanity/agent-context Studio plugin and add it to your Studio config:

import {defineConfig} from 'sanity'
import {structureTool} from 'sanity/structure'
import {agentContextPlugin} from '@sanity/agent-context/studio'

export default defineConfig({
  // ...
  plugins: [structureTool(), agentContextPlugin()],
})

Optionally, you can customize where the document type appears in the Studio structure using AGENT_CONTEXT_SCHEMA_TYPE_NAME:

// sanity.config.ts
import {defineConfig} from 'sanity'
import {structureTool} from 'sanity/structure'
import {agentContextPlugin, AGENT_CONTEXT_SCHEMA_TYPE_NAME} from '@sanity/agent-context/studio'

export default defineConfig({
	// ...
  plugins: [
    structureTool({
      structure: (S) =>
        S.list()
          .title('Content')
          .items([
            // Filter out agent context document from the default list
            ...S.documentTypeListItems().filter(
              (item) => item.getId() !== AGENT_CONTEXT_SCHEMA_TYPE_NAME,
            ),
            // Add it elsewhere, e.g. after a divider
            S.divider(),
            S.documentTypeListItem(AGENT_CONTEXT_SCHEMA_TYPE_NAME),
          ]),
    }),
    agentContextPlugin(),
  ],
})

• Create a new Agent Context document and copy the MCP URL from the top of the form.
• Connect to the MCP URL using any MCP-compatible client, authenticated with your Sanity read token (see Prerequisites). Here's an example using Vercel AI SDK:

import {createMCPClient} from '@ai-sdk/mcp'

 const mcpClient = await createMCPClient({
   transport: {
     type: 'http',
     url: '<CONTEXT_MCP_URL>',
     headers: {
       Authorization: `Bearer <SANITY_API_READ_TOKEN>`,
     },
   },
 })

• Verify it works by listing the available tools:

const tools = await mcpClient.tools()
 console.log(tools)

Agent Context MCP uses Sanity API tokens for authentication.

• Tokens are passed as Bearer tokens in the Authorization header
• Use a read token - agents only need read access
• Keep tokens server-side in your agent implementation – never expose them client-side

Agent Context MCP is read-only. Agents can query and retrieve content, but cannot create, update, or delete documents. What agents can access:

• Schema - document types, field definitions, and structure
• Content - published documents in the dataset
• References - agents can follow references between documents

By default, agents only see published documents. Set perspective to drafts to include draft content. The groqFilter field in the Agent Context document controls what content is visible. For example:

• _type == "product" - only product documents
• _type in ["article", "author"] - only articles and authors
• _type == "product" && public == true - only public products

If no filter is set, the agent can access all published documents the token has permission to read.

Your Sanity API token is missing or invalid. Ensure:

• The token has read access to the project
• You're passing it as a Bearer token in the Authorization header

Agent Context requires your schema to be available server-side. This happens automatically when your Studio runs, but if it's not working:

• Ensure you're on Sanity Studio v5.1.0+.
• Open your Studio in a browser - this triggers schema deployment
• Retry the MCP connection

• Check that the MCP URL is correct (project ID, dataset, slug)
• Verify the Agent Context document exists and is published
• Try calling tools() on your MCP client and log the result

Semantic search requires embeddings to be enabled for your dataset. If queries using text::semanticSimilarity aren't returning ranked results:

• Verify embeddings are enabled for your dataset
• Try explicitly adding ?embeddings=true to the MCP URL to rule out auto-detection issues