Reference

MCP Server

14 tools and 2 resources for AI agent access to the SourcePress knowledge graph, content pipeline, and eval loop.

MCP Server

The @sourcepress/mcp package exposes SourcePress capabilities to AI agents and MCP-compatible clients. It provides 14 tools and 2 resources covering the knowledge graph, content generation, eval loop, and approval pipeline.

Overview

The MCP server wraps the same pipeline available via the REST API into the Model Context Protocol interface. Agents connect to it and can query the knowledge graph, trigger generation, run evals, and submit content for approval — without direct HTTP calls.

Package: @sourcepress/mcp Protocol: @modelcontextprotocol/sdk Tools: 14 Resources: 2

Starting the Server

The MCP server runs as a standalone process. Start it via the CLI or directly:

# Via CLI (recommended)
pnpm sourcepress dev

# Direct invocation
node packages/mcp/dist/index.js

The server communicates over stdio by default, which is the standard transport for MCP clients (e.g., Claude Desktop, Cursor).

Configuration

The MCP server reads from the same sourcepress.config.ts used by the rest of the system. No separate config file is required.

// sourcepress.config.ts
import { defineConfig } from "@sourcepress/core";

export default defineConfig({
  knowledge: {
    dir: "./knowledge",
  },
  ai: {
    model: "claude-sonnet-4-5",
    budget: {
      daily: 10_00, // cents
    },
  },
  evals: {
    threshold: 0.8,
    max_iterations: 3,
  },
  github: {
    owner: "your-org",
    repo: "your-repo",
    branch: "main",
  },
});

Field	Type	Description
`knowledge.dir`	`string`	Directory scanned for knowledge files
`ai.model`	`string`	Anthropic model used for generation and judging
`ai.budget.daily`	`number`	Daily spend cap in cents
`evals.threshold`	`number`	Minimum score (0–1) to pass the eval loop
`evals.max_iterations`	`number`	Maximum generate-judge-improve cycles
`github.owner`	`string`	Repository owner for PR approval
`github.repo`	`string`	Repository name
`github.branch`	`string`	Target branch for approved content

Tools

The 14 tools are grouped by pipeline stage.

Knowledge Graph

Tool	Description
`knowledge_ingest`	Ingest a file or URL into the knowledge graph
`knowledge_query`	Query entities and relations from the graph
`knowledge_gaps`	Identify gaps in the knowledge graph
`knowledge_stale`	Detect stale knowledge entries

Content

Tool	Description
`content_generate`	Generate content against the knowledge graph
`content_improve`	Improve existing content using the prompt improvement function
`content_list`	List content items in a collection
`content_get`	Retrieve a single content item by slug

Eval

Tool	Description
`eval_run`	Run the generate-judge-improve loop for a given intent
`eval_score`	Score a piece of content against the knowledge graph
`eval_judge`	Run the judge function and return structured feedback

Approval

Tool	Description
`approval_submit`	Submit content for approval via GitHub PR
`approval_status`	Check the status of a pending approval PR

Schema

Tool	Description
`schema_get`	Return the current collection schema

Resources

The MCP server exposes 2 resources that agents can read directly.

Resource URI	Description
`sourcepress://graph`	Full serialized knowledge graph (entities, relations, clusters)
`sourcepress://schema`	Current collection and field schema

Resources are read-only. Mutations go through tools.

Tool Examples

`knowledge_ingest`

Ingests a local file or remote URL. Runs classify and entity extraction.

Input schema:

{
  "source": "https://example.com/article",
  "type": "url"
}

Response:

{
  "status": "ingested",
  "source": "https://example.com/article",
  "entities_found": 7,
  "relations_found": 4,
  "clusters": 1
}

`knowledge_query`

Queries the graph for entities matching a filter.

Input schema:

{
  "entity_type": "product",
  "limit": 10
}

Response:

{
  "entities": [
    {
      "id": "ent_01",
      "type": "product",
      "label": "SourcePress MCP Server",
      "relations": ["ent_02", "ent_05"]
    }
  ],
  "total": 1
}

`eval_run`

Runs the full generate-judge-improve loop for a given intent slug.

Input schema:

{
  "intent": "mcp-server-overview",
  "collection": "docs"
}

Response:

{
  "status": "passed",
  "iterations": 2,
  "final_score": 0.87,
  "content_slug": "mcp-server-overview",
  "threshold": 0.8
}

If the loop exhausts max_iterations without reaching threshold:

{
  "status": "failed",
  "iterations": 3,
  "final_score": 0.71,
  "threshold": 0.8,
  "reason": "Score did not reach threshold after 3 iterations."
}

`approval_submit`

Submits a content item for approval by opening a GitHub PR with a provenance block.

Input schema:

{
  "slug": "mcp-server-overview",
  "collection": "docs"
}

Response:

{
  "status": "submitted",
  "pr_url": "https://github.com/your-org/your-repo/pull/42",
  "pr_number": 42,
  "branch": "sourcepress/docs-mcp-server-overview"
}

`schema_get`

Returns the active collection schema.

Input schema: (no parameters)

Response:

{
  "collections": [
    {
      "name": "docs",
      "fields": [
        { "name": "title", "type": "string", "required": true },
        { "name": "description", "type": "string", "required": true },
        { "name": "section", "type": "string", "required": true },
        { "name": "order", "type": "number", "required": true }
      ]
    }
  ]
}

Connecting a Client

Claude Desktop

Add the server to your claude_desktop_config.json:

{
  "mcpServers": {
    "sourcepress": {
      "command": "node",
      "args": ["/path/to/your/project/packages/mcp/dist/index.js"],
      "env": {
        "SOURCEPRESS_API_KEY": "your-api-key",
        "GITHUB_TOKEN": "your-github-token"
      }
    }
  }
}

Cursor

Add to .cursor/mcp.json in your project root:

{
  "mcpServers": {
    "sourcepress": {
      "command": "node",
      "args": ["./packages/mcp/dist/index.js"]
    }
  }
}

Security

Write tools (knowledge_ingest, eval_run, approval_submit, content_generate, content_improve) require a valid SOURCEPRESS_API_KEY environment variable.
Read tools (knowledge_query, content_get, schema_get) and resources are unauthenticated by default in local development.
The MCP server inherits the CORS and rate-limiting posture of @sourcepress/server when run behind the REST layer.
AI and ingestion tools are rate-limited. Batch inputs are capped at 50 URLs per call.

REST API — same capabilities over HTTP
Knowledge Graph — entity and relation model
Eval Loop — generate-judge-improve internals
Approval Pipeline — GitHub PR provenance flow

MCP Server

Overview

Starting the Server

Configuration

Tools

Knowledge Graph

Content

Eval

Approval

Schema

Resources

Tool Examples

knowledge_ingest

knowledge_query

eval_run

approval_submit

schema_get

Connecting a Client

Claude Desktop

Cursor

Security

Related

`knowledge_ingest`

`knowledge_query`

`eval_run`

`approval_submit`

`schema_get`