Glooker

Developer impact analytics for GitHub organizations. Glooker pulls commit history, runs LLM-based analysis on each commit, and produces ranked developer reports with metrics like code complexity, PR discipline, AI-assisted coding percentage, and overall impact scores.

Quick Start

git clone https://github.com/Smartling/glooker.git
cd glooker
npm install
cp .env.example .env.local

Edit .env.local with your GitHub token and LLM API key, then:

npm run dev
# Open http://localhost:3000

That's it. Glooker uses SQLite by default — no database setup needed.

Features

• Jira integration — optionally tracks resolved Jira issues per developer, auto-discovers GitHub→Jira user mappings via commit emails
• Full commit coverage — uses GitHub's commit search API to capture all commits, not just PR-linked ones
• LLM-powered analysis — each commit is analyzed for complexity (1-10), type (feature/bug/refactor/etc), risk level, and whether it appears AI-generated
• AI detection — three layers: confirmed via Co-Authored-By trailers, PR body patterns (e.g. "Generated with Claude Code"), branch commit trailer scanning for merge commits, and LLM heuristic analysis
• PR discipline tracking — shows what percentage of each developer's commits went through pull requests
• Developer detail page — click any developer to see percentile rankings (vs avg/p50/p95), type breakdown, active repos, and full commit history with links to GitHub
• Progressive UI — developer table populates during report generation as each member completes, not after everything finishes
• Resumable reports — interrupted reports can be resumed, skipping already-analyzed commits (commit analyses save to DB inline)
• Scheduled reports — configure recurring reports on a cron schedule
• Export — CSV download, Google Sheets, or Download PDF (print-optimized layout)
• Multiple LLM providers — OpenAI, Anthropic, AWS Bedrock, any OpenAI-compatible endpoint (Ollama, vLLM, Azure), or Smartling AI Proxy

Report Metrics

Metric	Description
Commits	Total commits in period (all, not just PR-linked)
PRs	Merged pull requests authored
Lines +/-	Lines added and removed
Complexity	Mean LLM-assessed complexity (1-10)
PR%	Percentage of commits that went through a PR
AI%	Percentage of commits with AI assistance (confirmed + suspected)
Jira Issues	Resolved Jira tickets in period (optional, requires Jira config)
Reviews	Merged PRs reviewed by this developer
Impact	Weighted score: complexity (3.5) + PRs (2.7) + volume (2.0) + PR discipline (1.1) + Jira (0.5) + reviews (0.5) — max 10.3
Types	Commit categorization: feature, bug, refactor, infra, docs, test

Configuration

GitHub Token

Create a fine-grained personal access token:

• Resource owner: your org
• Repository access: All repositories
• Repository permissions: Contents (read), Pull requests (read), Metadata (read)
• Organization permissions: Members (read)

LLM Provider

Set LLM_PROVIDER in .env.local:

OpenAI (default)

LLM_PROVIDER=openai
LLM_API_KEY=sk-...
LLM_MODEL=gpt-4o

Anthropic

LLM_PROVIDER=anthropic
LLM_API_KEY=sk-ant-...
LLM_MODEL=claude-sonnet-4-20250514

Local / Custom (Ollama, vLLM, etc.)

LLM_PROVIDER=openai-compatible
LLM_BASE_URL=http://localhost:11434/v1
LLM_MODEL=llama3
LLM_API_KEY=not-needed

Apple Silicon (M1/M2/M3/M4) GPU acceleration: If you install Ollama via an x86 Homebrew (/usr/local/bin/brew), it will run under Rosetta and cannot access the Metal GPU — inference will be CPU-only and much slower. Install Ollama using the official installer or an ARM-native Homebrew (/opt/homebrew/bin/brew) to get full GPU acceleration. Verify with ollama ps — the PROCESSOR column should show 100% GPU, not 100% CPU.

AWS Bedrock

For AWS users with Bedrock model access enabled:

LLM_PROVIDER=bedrock
# Uses the default AWS credential provider chain (AWS_PROFILE, IAM role, etc.)
# AWS_PROFILE=your-profile
# AWS_REGION=us-east-1
LLM_MODEL=us.anthropic.claude-sonnet-4-6

For SSO authentication, run your dev server through aws-okta or aws sso login:

aws-okta exec your-profile -- npm run dev
# or
aws sso login --profile your-profile && AWS_PROFILE=your-profile npm run dev

Smartling AI Proxy

For Smartling customers with AI Proxy access:

LLM_PROVIDER=smartling
SMARTLING_BASE_URL=https://api.smartling.com
SMARTLING_ACCOUNT_UID=your_account_uid
SMARTLING_USER_IDENTIFIER=your_user_identifier
SMARTLING_USER_SECRET=your_user_secret
LLM_MODEL=anthropic/claude-sonnet-4-20250514

Jira (optional)

Track resolved Jira issues per developer. Requires a Jira Cloud or Server instance:

JIRA_ENABLED=true
JIRA_HOST=mycompany.atlassian.net
JIRA_USERNAME=your-email@company.com
JIRA_API_TOKEN=your-jira-api-token
JIRA_API_VERSION=3          # 3 for Cloud, 2 for Server
# JIRA_PROJECTS=PROJ1,PROJ2 # optional filter, default: all projects
# JIRA_STORY_POINTS_FIELDS=customfield_10016  # optional, instance-specific

GitHub→Jira user mappings are auto-discovered via commit author emails and cached in the user_mappings table. Editable in Settings > App Settings.

To find story points field IDs for your instance: GET https://<JIRA_HOST>/rest/api/3/field — look for fields with "story" or "point" in the name and use the id value (e.g. customfield_10016). If unset, story points are not collected.

Database

SQLite (default) — zero config, data stored in ./glooker.db:

# No config needed — this is the default

MySQL — for teams or production:

DB_TYPE=mysql
DB_HOST=localhost
DB_USER=root
DB_PASSWORD=
DB_NAME=glooker

Then initialize: mysql -u root < schema.sql

Prompt Customization

LLM prompts are stored as template files in the prompts/ directory. You can customize prompts by editing these files or pointing to a different directory:

PROMPTS_DIR=./my-custom-prompts

Template files use {{PLACEHOLDER}} syntax for dynamic values injected at runtime.

Each LLM-powered service has configurable temperature and max_tokens settings:

Setting	Default	Description
ANALYZER_TEMPERATURE	0	Commit analysis (deterministic)
ANALYZER_MAX_TOKENS	256	Commit analysis response limit
CHAT_AGENT_TEMPERATURE	0.3	Chat assistant
CHAT_AGENT_MAX_TOKENS	1500	Chat assistant response limit
CHAT_AGENT_MAX_ITERATIONS	5	Max tool-call rounds per chat
SUMMARY_TEMPERATURE	0.7	Developer summaries
SUMMARY_MAX_TOKENS	512	Developer summary response limit
HIGHLIGHTS_TEMPERATURE	0.5	Report comparison highlights
HIGHLIGHTS_MAX_TOKENS	512	Highlights response limit

All settings are optional — defaults match the original hardcoded values.

After editing prompt templates: Run npm test -- -u to update Jest snapshots, then review the diff to confirm the change is intentional. Snapshots assert the exact prompt text to prevent accidental regressions.

Docker

# Set your env vars
export GITHUB_TOKEN=github_pat_...
export LLM_API_KEY=sk-...

# Build and start (MySQL + app)
docker compose up --build -d

# Open http://localhost:3000

Note: The MySQL container exposes port 3307 by default (to avoid conflicts with a local MySQL on 3306). Edit docker-compose.yml to change this.

MCP Server

Glooker exposes a read-only MCP server at POST /api/mcp so you can query its data and analysis from Claude Code / Claude.ai — including cross-report analysis over time. It speaks stateless Streamable-HTTP JSON-RPC and provides 14 tools:

• Discovery: list_reports, get_org_summary
• Raw entities: query_commits, query_jira_issues, query_developer_stats, query_unmerged_work
• LLM/semantic: get_project_insights, get_project_details, get_highlights, get_team_pulse, get_developer_summary, get_release_notes, get_epic_summaries
• Time-series: get_metric_timeseries (metric × group_by = week/month/report/developer/repo/type)

Connect from Claude Code:

# Local dev server (see .mcp.json)
claude mcp add --transport http glooker-local http://localhost:3000/api/mcp

# Deployed (behind the Okta proxy)
claude mcp add --transport http glooker https://glooker-mcp.internal-tools.dev.smartling.net/mcp

Auth: primary auth is the mcp-okta-proxy sidecar (validates the Okta token, injects the user identity). The app also has a fail-closed backstop: when AUTH_ENABLED=true, /api/mcp returns 401 unless the proxy's identity header is present. Locally (with AUTH_ENABLED unset) the endpoint is open for direct connection. See .env.example for the trust model. All tools are read-only — no tool mutates data or triggers report runs.

Architecture

Browser (Next.js)  →  API Routes  →  GitHub API
                                  →  LLM Provider (OpenAI / Anthropic / Bedrock / Smartling / custom)
                                  →  Jira API (optional)
                                  →  SQLite or MySQL

How a report runs

1. List org members via GitHub API
2. For each member (pipelined — LLM starts while more members are still being fetched):
   • Search all commits and merged PRs in the date range
   • Fetch diffs for each commit
   • Detect AI co-authorship from commit trailers, PR body, and branch commits
   • Queue LLM analysis (concurrency-limited) for complexity, type, risk, and AI-generation detection
   • Save each commit analysis to DB immediately (enables resume)
3. When all of a member's commits are analyzed, aggregate and save developer stats to DB (enables progressive UI)
4. If Jira enabled: resolve each member's Jira account (via commit emails → user_mappings), fetch resolved issues via JQL, save to jira_issues
5. Final cross-member aggregation overwrites with canonical stats
6. Display ranked table with export options

Key files

src/lib/
├── llm-provider.ts          # LLM provider factory (OpenAI SDK for all providers)
├── bedrock-adapter.ts       # AWS Bedrock adapter (only loaded when provider=bedrock)
├── smartling-auth.ts         # Smartling OAuth (only loaded when provider=smartling)
├── github.ts                 # GitHub API: members, commit/PR search, diffs, AI detection
├── analyzer.ts               # LLM commit analysis prompt + response parsing
├── aggregator.ts             # Per-developer metric rollup
├── report-runner.ts          # Pipeline orchestrator (pipelined fetch → analyze → save)
├── jira/
│   ├── client.ts             # Jira REST API client (direct fetch, no external deps)
│   ├── mapper.ts             # GitHub→Jira user mapping (auto-discover + persist)
│   └── index.ts              # Re-exports
├── progress-store.ts         # In-memory progress tracking (developer-based)
├── schedule-manager.ts       # Cron-based scheduled report execution
├── schedule-validation.ts    # Schedule input validation
└── db/
    ├── index.ts              # DB abstraction (selects SQLite or MySQL)
    ├── sqlite.ts             # SQLite implementation (default)
    └── mysql.ts              # MySQL implementation

src/app/
├── page.tsx                  # Main dashboard (report list, generation, developer table)
├── report/[id]/dev/[login]/
│   └── page.tsx              # Developer detail page (percentiles, commits)
└── api/
    ├── report/[id]/dev/[login]/route.ts  # Developer detail API
    ├── report/[id]/commits/route.ts      # Commits per developer API
    └── ...                               # Other report & schedule endpoints

Development

npm run dev            # Start dev server
npm run build          # Production build
npm test               # Run all tests
npm run test:watch     # Run tests in watch mode
npm run test:coverage  # Run tests with coverage report

Tests live in src/lib/__tests__/ (unit and integration). CI runs tests automatically on every pull request via GitHub Actions.

If you see Cannot find module './638.js', run rm -rf .next and restart.

See CONTRIBUTING.md for contribution guidelines and CLAUDE.md for AI-assisted development context.

License

MIT