Massive Context MCP Server

Massive Context MCP

Handle massive contexts (10M+ tokens) with chunking, sub-queries, and free local inference via Ollama.

flowchart TD
    A[Claude Code] --> B[RLM MCP Server]
    B --> C{rlm_ollama_status}
    C -->|cached 60s| D{provider = auto}

    D -->|Ollama running| E[🦙 Ollama<br/>gemma3:12b]
    D -->|Ollama unavailable| F[☁️ Claude SDK<br/>claude-haiku-4-5]

    E --> G[["💰 $0<br/>Free local inference"]]
    F --> H[["💰 ~$0.80/1M<br/>Cloud inference"]]

    style A fill:#ff922b,color:#fff
    style B fill:#339af0,color:#fff
    style E fill:#51cf66,color:#fff
    style F fill:#748ffc,color:#fff
    style G fill:#51cf66,color:#fff
    style H fill:#748ffc,color:#fff

Based on the Recursive Language Model pattern. Inspired by richardwhiteii/rlm.

Core Idea

Instead of feeding massive contexts directly into the LLM:

1. Load context as external variable (stays out of prompt)
2. Inspect structure programmatically
3. Chunk strategically (lines, chars, or paragraphs)
4. Sub-query recursively on chunks
5. Aggregate results for final synthesis

Quick Start

Installation

Option 1: PyPI (Recommended)

uvx massive-context-mcp
# or
pip install massive-context-mcp

With Optional Extras:

# With Code Firewall integration (security filter for rlm_exec)
pip install massive-context-mcp[firewall]

# With Claude Agent SDK (for programmatic Claude API access)
pip install massive-context-mcp[claude]

# With all extras
pip install massive-context-mcp[firewall,claude]

Option 2: Claude Desktop One-Click

Download the .mcpb from Releases and double-click to install.

Option 3: From Source

git clone https://github.com/egoughnour/massive-context-mcp.git
cd massive-context-mcp
uv sync

Wire to Claude Code / Claude Desktop

Add to ~/.claude/.mcp.json (Claude Code) or claude_desktop_config.json (Claude Desktop):

{
  "mcpServers": {
    "massive-context": {
      "command": "uvx",
      "args": ["massive-context-mcp"],
      "env": {
        "RLM_DATA_DIR": "~/.rlm-data",
        "OLLAMA_URL": "http://localhost:11434"
      }
    }
  }
}

Tools

Setup & Status Tools

Analysis Tools

Quick Analysis with rlm_auto_analyze

For most use cases, just use rlm_auto_analyze — it handles everything automatically:

rlm_auto_analyze(
    name="my_file",
    content=file_content,
    goal="find_bugs"  # or: summarize, extract_structure, security_audit, answer:<question>
)

What it does automatically:

1. Detects content type (Python, JSON, Markdown, logs, prose, code)
2. Selects optimal chunking strategy
3. Adapts the query for the content type
4. Runs parallel sub-queries
5. Returns aggregated results

Supported goals:

Programmatic Analysis with rlm_exec

For deterministic pattern matching and data extraction, use rlm_exec to run Python code directly against a loaded context. This is closer to the paper's REPL approach and provides full control over analysis logic.

Tool: rlm_exec

Purpose: Execute arbitrary Python code against a loaded context in a sandboxed subprocess.

Parameters:

• code (required): Python code to execute. Set the result variable to capture output.
• context_name (required): Name of a previously loaded context.
• timeout (optional, default 30): Maximum execution time in seconds.

Features:

• Context available as read-only context variable
• Pre-imported modules: re, json, collections
• Subprocess isolation (won't crash the server)
• Timeout enforcement
• Works on any system with Python (no Docker needed)

Example — Finding patterns in a loaded context:

# After loading a context
rlm_exec(
    code="""
import re
amounts = re.findall(r'\$[\d,]+', context)
result = {'count': len(amounts), 'sample': amounts[:5]}
""",
    context_name="bill"
)

Example Response:

{
  "result": {
    "count": 1247,
    "sample": ["$500", "$1,000", "$250,000", "$100,000", "$50"]
  },
  "stdout": "",
  "stderr": "",
  "return_code": 0,
  "timed_out": false
}

Example — Extracting structured data:

rlm_exec(
    code="""
import re
import json

# Find all email addresses
emails = re.findall(r'\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b', context)

# Count by domain
from collections import Counter
domains = [e.split('@')[1] for e in emails]
domain_counts = Counter(domains)

result = {
    'total_emails': len(emails),
    'unique_domains': len(domain_counts),
    'top_domains': domain_counts.most_common(5)
}
""",
    context_name="dataset",
    timeout=60
)

When to use rlm_exec vs rlm_sub_query:

Tip: For large contexts, combine both — use rlm_exec to filter/extract, then rlm_sub_query for semantic analysis of filtered results.

Code Firewall Integration (Optional)

For enhanced security, integrate code-firewall-mcp to filter dangerous code patterns before execution:

pip install massive-context-mcp[firewall]

When installed, rlm_exec can automatically check code against a blacklist of known dangerous patterns (e.g., os.system(), eval(), subprocess with shell=True). The firewall uses structural similarity matching — normalizing code to its skeleton and comparing against blacklisted patterns via embeddings.

How it works:

1. Code is parsed to a syntax tree and normalized (identifiers → _, strings → "S")
2. Normalized structure is embedded via Ollama
3. Similarity is checked against blacklisted patterns in ChromaDB
4. Code is blocked if similarity exceeds threshold (default: 0.85)

Configuration (environment variables):

• RLM_FIREWALL_ENABLED=true — Enable firewall checks (auto-enabled when package installed)
• RLM_FIREWALL_MODE=warn|block — Warn or block on matches (default: warn)

Example blocked patterns:

• os.system(user_input) — Command injection
• eval(untrusted_data) — Code injection
• subprocess.Popen(..., shell=True) — Shell injection

Use rlm_firewall_status to check firewall availability and configuration.

Providers & Auto-Detection

RLM automatically detects and uses the best available provider:

How Auto-Detection Works

When you use provider="auto" (the default), RLM:

1. Checks if Ollama is running at OLLAMA_URL (default: http://localhost:11434)
2. Checks if gemma3:12b is available (or any gemma3 variant)
3. Uses Ollama if available, otherwise falls back to Claude SDK

The status is cached for 60 seconds to avoid repeated network checks.

Check Ollama Status

Use rlm_ollama_status to see what's available:

rlm_ollama_status()

Response when Ollama is ready:

{
  "running": true,
  "models": ["gemma3:12b", "llama3:8b"],
  "default_model_available": true,
  "best_provider": "ollama",
  "recommendation": "Ollama is ready! Sub-queries will use free local inference by default."
}

Response when Ollama is not available:

{
  "running": false,
  "error": "connection_refused",
  "best_provider": "claude-sdk",
  "recommendation": "Ollama not available. Sub-queries will use Claude API. To enable free local inference, install Ollama and run: ollama serve"
}

Transparent Provider Selection

All sub-query responses include which provider was actually used:

{
  "provider": "ollama",
  "model": "gemma3:12b",
  "requested_provider": "auto",
  "response": "..."
}

Autonomous Usage

Enable Claude to use RLM tools automatically without manual invocation:

1. CLAUDE.md Integration Copy CLAUDE.md.example content to your project's CLAUDE.md (or ~/.claude/CLAUDE.md for global) to teach Claude when to reach for RLM tools automatically.

2. Hook Installation Copy the .claude/hooks/ directory to your project to auto-suggest RLM when reading files >10KB:

cp -r .claude/hooks/ /Users/your_username/your-project/.claude/hooks/

The hook provides guidance but doesn't block reads.

3. Skill Reference Copy the .claude/skills/ directory for comprehensive RLM guidance:

cp -r .claude/skills/ /Users/your_username/your-project/.claude/skills/

With these in place, Claude will autonomously detect when to use RLM instead of reading large files directly into context.

Setting Up Ollama (Free Local Inference)

RLM can automatically install and configure Ollama on macOS with Apple Silicon. There are two installation methods with different trade-offs:

Choosing an Installation Method

Recommendation:

• Use Homebrew method if you have Homebrew and want managed updates
• Use Direct Download for automation, locked-down machines, or when you don't have admin access

Method 1: Homebrew Installation (Recommended if you have Homebrew)

# 1. Check if your system meets requirements
rlm_system_check()

# 2. Install via Homebrew
rlm_setup_ollama(install=True, start_service=True, pull_model=True)

What this does:

• Installs Ollama via Homebrew (brew install ollama)
• Starts Ollama as a managed background service (brew services start ollama)
• Pulls gemma3:12b model (~8GB download)

Requirements:

• macOS with Apple Silicon (M1/M2/M3/M4)
• 16GB+ RAM (gemma3:12b needs ~8GB to run)
• Homebrew installed

Method 2: Direct Download (Fully Headless, No Sudo)

# 1. Check system (Homebrew NOT required for this method)
rlm_system_check()

# 2. Install via direct download - no sudo, no Homebrew
rlm_setup_ollama_direct(install=True, start_service=True, pull_model=True)

What this does:

• Downloads Ollama from https://ollama.com/download/Ollama-darwin.zip
• Extracts to ~/Applications/Ollama.app (user directory, no admin needed)
• Starts Ollama via ollama serve (background process)
• Pulls gemma3:12b model

Requirements:

• macOS with Apple Silicon (M1/M2/M3/M4)
• 16GB+ RAM
• No special permissions needed!

Note on PATH: After direct installation, the CLI is at:

~/Applications/Ollama.app/Contents/Resources/ollama

Add to your shell config if needed:

export PATH="$HOME/Applications/Ollama.app/Contents/Resources:$PATH"

For Systems with Less RAM

Use a smaller model on either installation method:

rlm_setup_ollama(install=True, start_service=True, pull_model=True, model="gemma3:4b")
# or
rlm_setup_ollama_direct(install=True, start_service=True, pull_model=True, model="gemma3:4b")

Manual Setup

If you prefer manual installation or are on a different platform:

1. Install Ollama from https://ollama.ai or via Homebrew:

   brew install ollama
   

2. Start the service:

   brew services start ollama
   # or: ollama serve
   

3. Pull the model:

   ollama pull gemma3:12b
   

4. Verify it's working:

   rlm_ollama_status()
   

Provider Selection

RLM automatically uses Ollama when available. You can also force a specific provider:

# Auto-detection (default) - uses Ollama if available
rlm_sub_query(query="Summarize", context_name="doc")

# Explicitly use Ollama
rlm_sub_query(query="Summarize", context_name="doc", provider="ollama")

# Explicitly use Claude SDK
rlm_sub_query(query="Summarize", context_name="doc", provider="claude-sdk")

Usage Example

Basic Pattern

# 0. (Optional) First-time setup on macOS - choose ONE method:

# Option A: Homebrew (if you have it)
rlm_system_check()
rlm_setup_ollama(install=True, start_service=True, pull_model=True)

# Option B: Direct download (no sudo, fully headless)
rlm_system_check()
rlm_setup_ollama_direct(install=True, start_service=True, pull_model=True)

# 0b. (Optional) Check if Ollama is available for free inference
rlm_ollama_status()

# 1. Load a large document
rlm_load_context(name="report", content=<large document>)

# 2. Inspect structure
rlm_inspect_context(name="report", preview_chars=500)

# 3. Chunk into manageable pieces
rlm_chunk_context(name="report", strategy="paragraphs", size=1)

# 4. Sub-query chunks in parallel (auto-uses Ollama if available)
rlm_sub_query_batch(
    query="What is the main topic? Reply in one sentence.",
    context_name="report",
    chunk_indices=[0, 1, 2, 3],
    concurrency=4
)

# 5. Store results for aggregation
rlm_store_result(name="topics", result=<response>)

# 6. Retrieve all results
rlm_get_results(name="topics")

Processing a 2MB Document

Tested with H.R.1 Bill (2MB):

# Load
rlm_load_context(name="bill", content=<2MB XML>)

# Chunk into 40 pieces (50K chars each)
rlm_chunk_context(name="bill", strategy="chars", size=50000)

# Sample 8 chunks (20%) with parallel queries
# (auto-uses Ollama if running, otherwise Claude SDK)
rlm_sub_query_batch(
    query="What topics does this section cover?",
    context_name="bill",
    chunk_indices=[0, 5, 10, 15, 20, 25, 30, 35],
    concurrency=4
)

Result: Comprehensive topic extraction at $0 cost (with Ollama) or ~$0.02 (with Claude).

Analyzing War and Peace (3.3MB)

Literary analysis of Tolstoy's epic novel from Project Gutenberg:

# Download the text
curl -o war_and_peace.txt https://www.gutenberg.org/files/2600/2600-0.txt

# Load into RLM (3.3MB, 66K lines)
rlm_load_context(name="war_and_peace", content=open("war_and_peace.txt").read())

# Chunk by lines (1000 lines per chunk = 67 chunks)
rlm_chunk_context(name="war_and_peace", strategy="lines", size=1000)

# Sample 10 chunks evenly across the book (15% coverage)
sample_indices = [0, 7, 14, 21, 28, 35, 42, 49, 56, 63]

# Extract characters from each sampled section
rlm_sub_query_batch(
    query="List major characters in this section with brief descriptions.",
    context_name="war_and_peace",
    chunk_indices=sample_indices,
    provider="claude-sdk",  # Haiku 4.5
    concurrency=8
)

Result: Complete character arc across the novel — Pierre's journey from idealist to prisoner to husband, Natásha's growth, Nikolái Rostóv's journey from soldier to landowner — all for ~$0.03.

Data Storage

graph TD
    A[("$RLM_DATA_DIR")] --> B["📁 contexts/"]
    A --> C["📁 chunks/"]
    A --> D["📁 results/"]

    B --> B1[".txt files"]
    B --> B2[".meta.json"]
    C --> C1["by context name"]
    D --> D1[".jsonl files"]

    style A fill:#339af0,color:#fff
    style B fill:#51cf66,color:#fff
    style C fill:#51cf66,color:#fff
    style D fill:#51cf66,color:#fff

Contexts persist across sessions. Chunked contexts are cached for reuse.

Learning Prompts

Use these prompts with Claude Code to explore the codebase and learn RLM patterns. The code is the single source of truth.

Understanding the Tools

Read src/rlm_mcp_server.py and list all RLM tools with their parameters and purpose.

Explain the chunking strategies available in rlm_chunk_context.
When would I use each one?

What's the difference between rlm_sub_query and rlm_sub_query_batch?
Show me the implementation.

Understanding the Architecture

Read src/rlm_mcp_server.py and explain how contexts are stored and persisted.
Where does the data live?

How does the claude-sdk provider extract text from responses?
Walk me through _call_claude_sdk.

What happens when I call rlm_load_context? Trace the full flow.

Hands-On Learning

Load the README as a context, chunk it by paragraphs,
and run a sub-query on the first chunk to summarize it.

Show me how to process a large file in parallel using rlm_sub_query_batch.
Use a real example.

I have a 1MB log file. Walk me through the RLM pattern to extract all errors.

Extending RLM

Read the test file and explain what scenarios are covered.
What edge cases should I be aware of?

How would I add a new chunking strategy (e.g., by regex delimiter)?
Show me where to modify the code.

How would I add a new provider (e.g., OpenAI)?
What functions need to change?

License

MIT