MCP Integration

43 Tools. One Protocol. Your AI Agent's Document Library.

Connect any MCP-compatible AI agent to your local document corpus. Edge models drive accurate results without datacenter hardware. Only index computation runs locally — search and retrieval are instant.

Protocol

What Is the Model Context Protocol?

MCP is a standard interface that connects AI models to external tools and data sources. NeuronCite speaks MCP natively via stdin/stdout JSON-RPC 2.0, exposing 43 tools that any compatible agent can discover and invoke.

AI Agent

Claude Code, custom agents

MCP Server

neuroncite mcp serve

Your Documents

Local index + embeddings

Architecture

Edge Model Advantage

All computation runs on your hardware. Embedding and LLM inference happen locally, while search and retrieval are pure mathematical operations with zero GPU cost.

Local Embeddings

Small transformer models (384–4096 dimensions) compute document embeddings on your hardware. Eight models available from bge-small (33M params) to e5-large-v2 (335M params). CPU works fine; GPU accelerates batch indexing.

Ollama LLM Integration

Quantized language models run locally via Ollama for autonomous citation verification. Models like llama3.1:8b or mistral:7b provide strong reasoning on consumer hardware without cloud API keys.

Instant Retrieval

HNSW vector search and BM25 keyword matching are pure mathematical operations. Once documents are indexed, every search query returns results in milliseconds with zero GPU cost.

Workflow

Autonomous Citation Verification

A six-stage pipeline driven entirely by your AI agent. From LaTeX parsing to annotated PDF reports, every step executes through MCP tool calls.

Parse LaTeX

Extract all \cite{} commands from .tex files, resolve cite-keys against .bib entries.

Match PDFs

Match bibliographic entries to indexed PDFs via Jaro-Winkler filename similarity.

Create Batches

Group citations into batches for parallel sub-agent processing.

Verify Claims

Sub-agents search cited PDFs, run multi-query verification, produce verdicts.

Submit Results

Collect verdicts, confidence scores, passages, corrections, and flags.

Export Reports

Generate annotated PDFs, CSV/XLSX reports, JSON corrections, statistics.

Tools

43 MCP Tools

Each tool maps to a JSON-RPC method with typed parameters and structured responses. Organized into eight functional categories.

Session Management

neuroncite_sessions

List all sessions with metadata and statistics

neuroncite_session_delete

Delete session and all associated data

neuroncite_session_update

Update session label, metadata, or tags

neuroncite_session_diff

Compare file differences between two sessions

neuroncite_discover

Scan directory for supported files and indexing status

Indexing

neuroncite_index

Index documents from directory or cached HTML pages (PDF, HTML)

neuroncite_index_add

Incrementally add or update files in existing session

neuroncite_reindex_file

Re-extract, re-chunk, and re-embed a single file

neuroncite_preview_chunks

Preview chunking output before committing

Search & Retrieval

neuroncite_search

Hybrid vector + BM25 search with optional reranking

neuroncite_batch_search

Multiple queries in a single call

neuroncite_multi_search

Search across 2–10 sessions simultaneously

neuroncite_compare_search

Side-by-side results from two sessions

neuroncite_text_search

Literal substring search within indexed documents

neuroncite_content

Retrieve full page text by file ID and page number

neuroncite_batch_content

Batch retrieve content from multiple documents

neuroncite_export

Export search results as Markdown, BibTeX, CSL-JSON, RIS

File & Chunk Inspection

neuroncite_files

List indexed files with per-file statistics

neuroncite_chunks

Browse document chunks with pagination

neuroncite_file_compare

Compare files across different sessions

neuroncite_quality_report

Text extraction quality overview for a session

Citation Verification

neuroncite_citation_create

Create verification job from LaTeX and BibTeX files

neuroncite_citation_claim

Claim a batch for sub-agent verification

neuroncite_citation_submit

Submit verification results for a batch

neuroncite_citation_status

Get job status with verdict distribution

neuroncite_citation_rows

Retrieve individual citation rows with pagination

neuroncite_citation_export

Export results as CSV, XLSX, JSON, annotated PDFs

neuroncite_citation_retry

Reset failed or flagged rows for re-verification

neuroncite_citation_fetch_sources

Download cited PDFs and HTML pages from BibTeX URLs/DOIs

Annotation

neuroncite_annotate

Highlight text passages in PDFs with color-coded annotations

neuroncite_annotate_status

Per-quote progress for annotation jobs

neuroncite_inspect_annotations

Inspect highlight annotations in a PDF file

neuroncite_annotation_remove

Remove highlights by color, page, or all

Web Sources

neuroncite_html_fetch

Download and cache HTML pages with metadata extraction

neuroncite_html_crawl

Crawl websites via BFS link-following or sitemap discovery

neuroncite_bib_report

Generate CSV/XLSX report from BibTeX file entries

System & Models

neuroncite_models

List available embedding models with configuration details

neuroncite_reranker_load

Load cross-encoder reranker model at runtime

neuroncite_health

Server health status and build features

neuroncite_doctor

System capabilities check (GPU, CUDA, Tesseract, pdfium)

neuroncite_jobs

List all indexing and annotation jobs

neuroncite_job_status

Get detailed status of a specific job

neuroncite_job_cancel

Cancel a queued or running job

Setup

Three Ways to Start the MCP Server

Launch the MCP server from the CLI, register it automatically via the built-in install command, or start it from the native GUI. All paths lead to the same stdin/stdout JSON-RPC 2.0 server.

CLI

Run neuroncite mcp serve to start the stdio server directly. Pass --model to override the default embedding model.

Auto-Install

Run neuroncite mcp install to register the server in Claude Code's settings (~/.claude/settings.json). Reverse with neuroncite mcp uninstall.

Native GUI

Open the Settings tab, navigate to the MCP Server panel, and click Install. The GUI handles registration and displays the current config path.

                            ~/.claude/settings.json (Claude Code)
                        
{
  "mcpServers": {
    "neuroncite": {
      "command": "neuroncite",
      "args": ["mcp", "serve"]
    }
  }
}

How It Works

NeuronCite starts as an MCP server communicating via stdin/stdout JSON-RPC 2.0
The AI agent discovers available tools through the MCP protocol handshake
All indexing, search, and verification operations run locally on your machine
No API keys and no cloud services. Optional tools (HTML fetch, citation source download, DOI resolution) make outbound requests when explicitly invoked