Pi

Pi (source: earendil-works/pi) is a minimal, extensible terminal coding agent. Its unified LLM layer talks to any provider through ~/.pi/agent/models.json — including plain OpenAI Chat Completions endpoints — so GreenPT connects directly with no translation proxy.

What's already open, before you install anything

You don't need an API key, Pi, or a config file to read GreenPT's own onboarding material — some of it is served publicly, unauthenticated, straight off the wire:

curl -s https://api.greenpt.ai/v1/agents
curl -s https://api.greenpt.ai/v1/pricing
curl -s https://api.greenpt.ai/v1/skills/catalog
curl -s https://api.greenpt.ai/v1/skills/catalog/greenpt

That last one is worth pulling up first: it's a self-contained walkthrough you can read before deciding whether to set anything up. The same content is also injectable into a real conversation — any authenticated chat completion can add "skills": "greenpt" to its request body to get it applied as a system prompt, so an agent operating on your behalf can pull in this guidance without you pasting it in by hand.

None of this makes a completion for you — you still need a GreenPT API key for that — but it means the "how do I use this" step has no signup wall.

Prerequisites

• Pi installed:

  npm install -g --ignore-scripts @earendil-works/pi-coding-agent
  # or
  curl -fsSL https://pi.dev/install.sh | sh

• A GreenPT API key: manage API keys

1. Set your API key

echo 'export GREENPT_API_KEY="your-greenpt-key"' >> ~/.zshrc   # or ~/.bashrc
source ~/.zshrc

2. Add the GreenPT provider

Add a greenpt provider to Pi's custom model config, ~/.pi/agent/models.json:

{
  "providers": {
    "greenpt": {
      "baseUrl": "https://api.greenpt.ai/v1",
      "api": "openai-completions",
      "apiKey": "$GREENPT_API_KEY",
      "models": [
        { "id": "glm-5.3" },
        { "id": "deepseek-v4.1-flash-1017" }
      ]
    }
  }
}

This file reloads every time you open /model — edit it during a session, no restart needed.

Keep the $ on $GREENPT_API_KEY — without it, Pi reads the text as a literal key instead of an environment variable.

Recommended models

Any model from the catalogue works — add its id to the block above. See Model Cards for the full list and prices.

Optional: keep the model list in sync

greenpt-pi-sync.sh rewrites only .providers.greenpt, cross-checking our recommended coding models against what's live so a retired model drops out instead of leaving a dead entry behind. Re-run it whenever you want a refresh.

curl -fsSL https://docs.greenpt.ai/scripts/greenpt-pi-sync.sh -o greenpt-pi-sync.sh
shasum -a 256 greenpt-pi-sync.sh   # 0eefbddb07156e83d269f48fedfb2d6b01650b9276313b273dbaa39adaec5e4e
sh greenpt-pi-sync.sh

It writes to ~/.pi/agent/models.json using credentials from your shell, so a mismatched hash means the download was tampered with — don't run it. Set GREENPT_BASE_URL to sync against a different environment.

3. Select the model

pi --provider greenpt --model glm-5.3

Or start pi and run /model to choose interactively; greenpt appears as a provider with your configured models underneath.

4. Verify the connection

Quick standalone check of the same endpoint Pi will call:

curl -s -H "Authorization: Bearer $GREENPT_API_KEY" -H 'Content-Type: application/json' \
  "https://api.greenpt.ai/v1/chat/completions" \
  -d '{"model":"glm-5.3","messages":[{"role":"user","content":"Reply with exactly: OK"}],"max_tokens":64}' \
  | jq '.choices[0].message.content'

If this returns "OK", Pi will connect the same way — fix auth or the base URL here first before troubleshooting inside Pi.

Live status footer: green-meter

Pi's default footer squeezes everything into flat, run-on lines with no label/value distinction. green-meter replaces it with labeled, color-weighted groups, and — for the greenpt provider specifically — turns the usual $ cost line into GreenPT's real signals: live credit spend and environmental impact.

You're at: ~/Dev/ozan/studio-pi4 · on branch, main
Usage: ↑12k · ↓664 · Draw: €0.02/m · €1.30/h · €0.043 sess · Impact: 6.7 g/m · 160 g/h · 288.0 g sess · 1.49 kW
Session: started 12m ago · Context: 41.7%   Model: deepseek (off)

• Draw — €/min, €/hour, and cumulative this session, from the live credit balance (X-Credits-Remaining on GET /v1/agents).
• Impact — CO₂e (per-minute, per-hour, session total) and average kW power draw, from GET /v1/profile/impact.

Install it

Download and verify before placing the file — it's a Pi extension, which runs with the same trust as any other code you load into your terminal:

curl -fsSL https://docs.greenpt.ai/extensions/green-meter.ts -o green-meter.ts
shasum -a 256 green-meter.ts

The hash should be:

a998b478c1b0dddee625649003b528bce78eede110e390a07aa8d567cf2ab34f  green-meter.ts

Then place it where Pi auto-discovers extensions, and reload:

mkdir -p ~/.pi/agent/extensions
mv green-meter.ts ~/.pi/agent/extensions/green-meter.ts

/reload

It's enabled by default once loaded. Toggle it, or get an in-session summary, with:

/green-meter       # off → restore Pi's default footer, on → re-enable
/green-meter help  # what it shows and where the numbers come from

Caveats

• Whole-account figures. Both the credit balance and the impact endpoint are keyed to your account, not this Pi session — another process spending on the same key inflates Draw and Impact. They're session-relative signals, not per-run meters.
• Session-scoped, persisted across reloads. Values reset when you start a new session, but survive a /reload of the same one.
• GreenPT only. The Draw/Impact groups require the greenpt provider; other providers fall back to Pi's native $ cost line.

If you already have a GreenPT API key and just want this without reading the rest of this page, GET /v1/skills/catalog/greenpt (see above) carries the same install steps as agent-readable instructions — hand it to your own coding agent and let it set this up for you.

Troubleshooting