Nft Log Analyzer

AI-powered log analysis that automatically files GitHub Issues — 100% local via Ollama, zero data leaves your machine.

Python Ollama MCP License

---

What It Does

Point it at any log file and it will:

1. Scan 500MB+ files in seconds using ripgrep
2. Parse error patterns, deduplicate repeated events
3. Analyse using local LLM (Ollama + deepseek-r1:14b) via CrewAI agents
4. Compose structured GitHub Issues with root cause and suggested fixes
5. File Issues automatically to your repo — skipping duplicates

All processing happens locally on your machine. Raw log content never leaves your system.

---

Architecture

Claude Desktop / Cursor / LangChain
         ↓  MCP (stdio or HTTP+SSE)
   MCP Log Analyzer Server
         ↓
   ripgrep pre-filter (2-4s on 500MB)
         ↓
   mmap streaming parser + deduplicator
         ↓
   CrewAI agents → Ollama (local LLM)
         ↓
   GitHub Issues API

---

Requirements

Requirement	Version	Notes
Python	3.11+	3.14 not supported
Ollama	Latest	brew install ollama
deepseek-r1:14b	—	~9GB download
ripgrep	Latest	brew install ripgrep
RAM	16GB min	32GB recommended
macOS	Ventura 13+	Apple Silicon recommended

---

Quick Start

1. Install system dependencies

brew install ollama ripgrep
brew services start ollama
ollama pull deepseek-r1:14b   # ~9GB — start this first

2. Clone and set up Python environment

git clone https://github.com/YOUR_ORG/mcp-log-analyzer
cd mcp-log-analyzer

/opt/homebrew/bin/python3.11 -m venv .venv
source .venv/bin/activate

pip install --upgrade pip
pip install mcp "crewai>=0.80.0" crewai-tools langchain-ollama \
    litellm fastapi uvicorn httpx httpx-sse \
    structlog loguru pydantic python-dotenv \
    tenacity rich typer

3. Configure environment

cp .env.example .env
nano .env   # fill in your values

GITHUB_PAT=ghp_your_token_here
GITHUB_REPO_OWNER=your-username
GITHUB_REPO_NAME=your-repo
OLLAMA_BASE_URL=http://localhost:11434
OLLAMA_MODEL=deepseek-r1:14b
CREWAI_TELEMETRY_OPT_OUT=true
OTEL_SDK_DISABLED=true
OLLAMA_KEEP_ALIVE=-1

4. Create a GitHub PAT

Go to: github.com → Settings → Developer settings → Personal access tokens → Tokens (classic)

Enable scope: repo (full)

5. Register with Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "mcp-log-analyzer": {
      "command": "/path/to/mcp-log-analyzer/.venv/bin/python",
      "args": ["/path/to/mcp-log-analyzer/mcp_server/server.py"],
      "env": {
        "GITHUB_PAT": "ghp_your_token",
        "GITHUB_REPO_OWNER": "your-username",
        "GITHUB_REPO_NAME": "your-repo",
        "OLLAMA_BASE_URL": "http://localhost:11434",
        "OLLAMA_MODEL": "deepseek-r1:14b"
      }
    }
  }
}

Restart Claude Desktop. You should see the 🔨 tools icon appear.

---

Usage

Via Claude Desktop (natural language)

analyze the log file at /var/log/app.log and file GitHub issues for any errors

use analyze_log_file with path="/var/log/app.log" dry_run=true

check status of job abc12345

Via Python CLI

source .venv/bin/activate

python3 -c "
from dotenv import load_dotenv
load_dotenv()
from mcp_server.tools.analyze_tool import analyze_log_file
import asyncio, json

result = asyncio.run(analyze_log_file({
    'path': '/var/log/app.log',
    'severity': 'ERROR',
    'dry_run': False
}))
print(result[0].text)
"

---

MCP Tools Reference

ping

Health check — verifies the server and Ollama are running.

{}

Returns: "mcp-log-analyzer online — Ollama: deepseek-r1:14b"

---

analyze_log_file

Start async log analysis. Returns a job ID immediately — pipeline runs in background.

Parameter	Type	Required	Default	Description
path	string	✅	—	Absolute path to log file
severity	string	—	ERROR	Minimum severity: WARN, ERROR, CRITICAL
dry_run	boolean	—	false	Preview issues without filing to GitHub

Returns:

{
  "job_id": "abc12345",
  "status": "started",
  "message": "Analysis started. Check progress with get_job_status('abc12345')."
}

---

get_job_status

Check the status of a running analysis job.

Parameter	Type	Required	Description
job_id	string	✅	Job ID returned by analyze_log_file

Returns (running):

{
  "status": "running",
  "job_id": "abc12345",
  "lines_filtered": 487,
  "chunks": 1
}

Returns (done):

{
  "status": "done",
  "job_id": "abc12345",
  "lines_filtered": 487,
  "unique_events": 4,
  "chunks": 1,
  "issues_filed": 2,
  "github_issues": [
    {
      "title": "[CRITICAL][minting-service] DB connection pool exhausted (x117)",
      "url": "https://github.com/your-org/your-repo/issues/42",
      "number": 42
    }
  ]
}

---

Compatible MCP Clients

Client	Transport	Config
Claude Desktop	stdio	claude_desktop_config.json
Claude Code CLI	stdio	.mcp.json in project root
Cursor	stdio or HTTP+SSE	.cursor/mcp.json
LangChain	HTTP+SSE	url: http://localhost:8000/sse
n8n	HTTP+SSE	HTTP Request node → SSE

HTTP+SSE Transport (for Cursor, LangChain, n8n)

python mcp_server/server.py --transport sse --port 8000

---

Customising with Skills

Skills are plain English .md files that teach the agents your stack's error patterns. Three built-in skills ship with the project:

Skill	Purpose
skills/nft-app-errors.skill.md	NFT/blockchain error classification
skills/infrastructure-errors.skill.md	Infrastructure error classification
skills/bug-composition.skill.md	GitHub Issue format rules

Writing your own skill

Create skills/my-stack-errors.skill.md:

# My Stack Error Classification

## CRITICAL — file bug immediately
- "FATAL: database connection refused" = service down
- "out of memory" = process crash imminent

## HIGH — file bug, non-urgent  
- "connection timeout" on external API = degraded performance

## IGNORE — known false positives
- "reconnecting..." during deploys = expected

Then load it in agents/crew.py:

_load_skill("my-stack-errors.skill.md")

---

Pipeline Internals

500MB log file
    ↓  ripgrep (2-4 seconds)
    ↓  Filters: ERROR|FATAL|CRITICAL|WARN|Exception|Traceback
~5MB of error lines
    ↓  mmap streaming parser
    ↓  LogEvent objects with timestamp, level, component, message
    ↓  Deduplicator (fingerprints strip req_id, numbers, hex)
4-20 unique error patterns
    ↓  Chunker (10 events per chunk, CRITICAL first)
1-3 chunks
    ↓  Single CrewAI agent → Ollama (local)
    ↓  Structured bug reports in markdown
    ↓  Title extractor + label classifier
    ↓  Duplicate check via GitHub search API
GitHub Issues filed

---

Performance

Tested on Apple Silicon (M2, 32GB):

File size	Filter time	Analysis time	Total
10MB	<1s	3-5 min	~5 min
100MB	1-2s	3-5 min	~7 min
500MB	3-5s	5-10 min	~15 min

Analysis time depends on number of unique error patterns found (not file size).

---

Troubleshooting

Symptom	Fix
ollama ps shows empty	Run ollama run deepseek-r1:14b then /bye to warm the model
MCP server disconnected in Claude Desktop	Check ~/Library/Logs/Claude/mcp-server-*.log for Python errors
Issues filed: 0	Verify GITHUB_PAT in claude_desktop_config.json is a real token, not placeholder
Timeout after 600s	Add OLLAMA_KEEP_ALIVE=-1 to .env and restart Ollama
crewai install fails	Requires Python 3.11 — not compatible with 3.13/3.14
Permission denied on /usr/local/bin	Use /opt/homebrew/bin/ instead on Apple Silicon

---

Roadmap

v1 (current)

• Local filesystem log ingestion
• ripgrep + mmap pipeline
• Single-agent CrewAI analysis
• GitHub Issues filing with dedup
• Claude Desktop + stdio MCP transport

v2 (planned)

• Datadog MCP integration
• Splunk MCP integration
• HTTP+SSE transport (Cursor, LangChain, n8n)
• Scheduled analysis triggers
• Parallel chunk processing
• Web dashboard for job history

---

Contributing

Contributions welcome — especially new skill files for different stacks.

1. Fork the repo
2. Create skills/your-stack-errors.skill.md
3. Test it against a real log file
4. Open a PR with example output

---

License

MIT — see LICENSE