Project: Climate Awareness Nudging

Integration Points — Where to Nudge

Open WebUI

Open WebUI is a self-hosted frontend for LLMs. It has a plugin/pipeline system and customizable UI.

Location	Nudge type	Difficulty	Notes
Chat input area	Prompt complexity hint (leaf icon)	Low	Add icon next to send button that reacts to prompt length
Chat input area	"Do you need AI for this?" tooltip	Medium	Requires simple query classification
Sidebar / header	Session cost ticker	Low	Small text showing running estimate
New chat screen	Purpose selector tags	Low	Optional tags: Learning / Work / Exploring / Fun
After N messages	Conversation length banner	Low	Dismissable info banner
Response footer	Per-response cost estimate	Low	Small text below each AI response
Settings / Profile	Personal usage dashboard	Medium	New page with charts and equivalences
Admin panel	Org-wide usage dashboard	Medium	Aggregated, anonymized stats
Model selector	Model efficiency labels	Low	Show relative cost/efficiency per model

Pipelines (server-side) — intercept requests/responses, add metadata (token count, estimated cost), inject nudge logic
Custom JS/CSS injection — Open WebUI supports custom scripts; lightweight UI nudges can be added this way
Filter pipelines — analyze prompts before sending, can add pre-request nudges
Response metadata — attach cost estimates to responses, render in UI

Token counter pipeline — counts input/output tokens, converts to estimated energy/CO2, attaches as metadata
Simplicity detector pipeline — classifies if a query is "too simple for AI", suggests alternatives
Redundancy detector pipeline — compares new prompt to recent conversation history, flags repetition
Chain detector pipeline — detects "generate-then-summarize" patterns across org conversations

Claude Code is a terminal-based AI coding assistant. Nudging here must respect the CLI paradigm — concise, non-intrusive, text-based.

Location	Nudge type	Difficulty	Notes
Status line	Session token/cost counter	Low	Already has a status line — add cost estimate
After response	Per-response cost footnote	Low	Small line showing tokens used + estimate
Session end	Session summary	Low	"This session: X tokens, ~Yg CO2"
Periodic (every N messages)	Reflection prompt	Low	"Long session — would a fresh start be more efficient?"
Model selection	Efficiency suggestion	Medium	"This task might work with Haiku (10x cheaper/greener)"
Hooks (post-response)	Custom awareness hook	Low	Users can add shell hooks that display info after each response

Hooks system — Claude Code supports pre/post hooks; a post-response hook could display cost info
Custom status line — configure the status line to show running cost estimates
MCP server — build a small MCP server that tracks usage and provides cost data
Wrapper script — a thin shell wrapper that tracks and reports on Claude Code usage

Post-response hook — shell script that reads token count from the response and displays a one-line cost estimate
Session summary hook — on session end, display total usage and CO2 equivalent
Weekly digest — cron job that aggregates usage data and emails/displays a weekly summary
.claude/climate-config — user config for preferred equivalence units, nudge frequency, etc.

For organizations running their own AI gateway (like AarhusAI), a proxy layer can:

Log all requests/responses with token counts and model info
Inject response headers with cost metadata
Serve a dashboard aggregating usage across all frontends
Apply rate-based nudges — if a user exceeds a threshold, add a gentle note to the next response
Detect organizational patterns — like AI-generated-text-to-AI-summarize chains

Regardless of frontend, these components are reusable: