GreenPipe: AI CI/CD Carbon Analyzer

AI agent that measures pipeline carbon using ISO SCI standard, benchmarks against 200 real pipelines, and auto-creates optimization MRs reducing emissions by 35% with one comment.

Comment

Inspiration

CI/CD pipelines run millions of times daily. A 2025 study (arxiv:2510.26413) found GitHub Actions alone produced 456.9 megatons of CO₂ in 2024 — equivalent to 7,600 trees capturing carbon for a year. Our analysis of 200 real GitLab pipelines revealed that 76% have no caching, 88% lack interruptible flags, and the average pipeline wastes over 35% of its energy on redundant work.

A systematic mapping study of 92 papers (MDPI 2025) found zero research on AI-driven carbon optimization of CI/CD. Tools exist to measure (Eco-CI) or shift regions (CarbonRunner), but nothing automatically measures AND fixes the problem.

GreenPipe was born from this gap: what if an AI agent could analyze any pipeline, calculate its carbon footprint using a real ISO standard, and create a merge request with the fix — all from a single comment?

What it does

GreenPipe is a GitLab Duo AI agent that performs the complete green optimization loop:

Measure → Analyze → Fix → Validate → Commit → MR

  1. Triggered by mentioning @greenpipe on any issue
  2. Discovers project context dynamically (works on any GitLab project)
  3. Reads .gitlab-ci.yml, carbon budget config, and benchmark data
  4. Calculates SCI score using ISO/IEC 21031:2024: \( \text{SCI} = \frac{(E \times I) + M}{R} \)
  5. Detects 13 waste patterns (no cache, oversized images, redundant installs, etc.)
  6. Benchmarks against our dataset of 200 real GitLab pipelines
  7. Generates optimized .gitlab-ci.yml and validates with ci_linter
  8. Creates a branch, commits the fix, and opens a merge request with the full SCI report
  9. Accounts for its own carbon cost — 1.5 gCO₂e analysis vs 390 gCO₂e/month savings = 260× ROI

Key Features

  • ISO/IEC 21031:2024 SCI Standard — not arbitrary scores
  • Per-job carbon attribution — shows exactly which job wastes the most
  • 200-pipeline benchmark dataset — rates your pipeline vs real-world data
  • Carbon budget enforcement — configurable via .gitlab/greenpipe.yml
  • Carbon-aware scheduling — suggests low-carbon regions (e.g., "Move to europe-north1 for 97% reduction")
  • Agent self-cost transparency — reports its own inference carbon alongside savings

How I built it

Platform: GitLab Duo Agent Platform with a custom flow definition

AI Model: Claude (Anthropic) via GitLab Duo — powers the analysis, optimization, and report generation

Scientific Foundation:

  • ISO/IEC 21031:2024 SCI specification
  • SPECpower database for runner power coefficients (AMD EPYC 7B12: 7W TDP per job)
  • Cloud Carbon Footprint methodology for embodied carbon (34.25 gCO₂e/hour)
  • Google Cloud region carbon intensity data

Dataset: Built a collection of 200 real .gitlab-ci.yml files from public GitLab projects using the GitLab API. Analyzed each for SCI scores, waste patterns, and optimization potential. Key findings:

Metric Value
Average waste score 5.5 / 9
SCI Median 623 gCO₂e/run
SCI P90 2,144 gCO₂e/run
Average reduction potential 35.6%

Agent Tools Used:

  • get_project — dynamic project discovery
  • read_files — pipeline config + carbon budget + benchmarks
  • ci_linter — validates optimized YAML before committing
  • create_commit — commits fix on a timestamped branch
  • create_merge_request — opens MR with SCI analysis report
  • create_issue — creates tracking issue with full methodology

Challenges I ran into

Flow timeout: The initial prompt generated reports too large for the agent's output limits. I iteratively compressed the report template while maintaining scientific rigor — the optimized YAML goes into the commit, not the report text.

Service account permissions: The GitLab Duo service account couldn't post issue comments (create_issue_note). Pivoted to using create_issue for the report and create_commit + create_merge_request for the fix.

AI Catalog versioning: Discovered that flow updates require new Git tags to be picked up by the AI Catalog. Each iteration needed a tag bump (went through v1 to v1.7).

Dynamic project support: Removing hardcoded project IDs required using context:project_id as input and calling get_project to discover the default branch — couldn't assume "main".

Branch naming: Fixed branch names (greenpipe/optimize-pipeline) broke on the second run. Switched to timestamped branches (greenpipe/optimize-20260324-1200).

Accomplishments that I'm proud of

  • Working end-to-end auto-fix MR with validated, passing CI pipeline
  • ISO/IEC 21031:2024 compliance — the only green CI/CD tool using the actual international standard
  • 200-pipeline benchmark dataset — real data, not theoretical
  • Agent self-cost accounting — honest carbon accounting that no other tool provides
  • Tested on real-world pipelines — not toy examples
  • Works on any GitLab project — no hardcoded configuration needed

What I learned

  • 76% of real CI/CD pipelines have no caching — the waste is enormous
  • AI agents can perform complete Measure → Fix → Commit loops autonomously
  • The GitLab Duo Agent Platform is powerful but requires careful prompt engineering
  • Using an ISO standard instead of arbitrary scores gives real credibility
  • Accounting for the AI's own carbon cost is essential for honest green tools

What's next for GreenPipe

  • Multi-agent architecture — dedicated Collector, Analyzer, and Publisher agents
  • Real pipeline duration data — integrate with GitLab Pipeline API for actual measurements
  • CI/CD Component — one-line include: installation for any project
  • Carbon trend dashboard — track SCI scores over time via GitLab Pages
  • Water footprint — add water usage estimation alongside carbon

Built With

  • claude-(anthropic)
  • cloud-carbon-footprint
  • gitlab-api
  • gitlab-duo-agent-platform
  • iso/iec-21031:2024-sci-standard
  • python
  • specpower-database
Share this project:

Updates