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.

UI-level integration points

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

Technical approaches in Open WebUI

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

Specific pipeline ideas

• 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 (CLI)

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

Integration points

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

Technical approaches in Claude Code

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

Specific implementation ideas

• 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.

---

General (platform-agnostic)

API/proxy layer

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

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

For self-hosted inference where the gateway sits on the same machines as the GPUs, carbontracker is a natural fit for the measurement layer: it samples real hardware power draw and combines it with regional grid carbon intensity, replacing bottom-up token-based estimates with measured energy. Wrap inference calls with the Python library or run the CLI alongside the workload, then attach the resulting per-request energy/CO2 to the response headers proposed in CO2 Research § 5. For hosted models (OpenAI, Anthropic, etc.) carbontracker can't reach the hardware, so estimates remain the only option until providers expose telemetry themselves.

Shared components

Regardless of frontend, these components are reusable:

• Cost estimation library — token count → energy → CO2 conversion
• Equivalence formatter — "Xg CO2" → human-friendly comparison
• Usage storage — lightweight database for tracking per-user usage over time
• Nudge rule engine — configurable rules for when/how to nudge