Vinci in the Pi coding agent

Use Vinci inside Pi , the MIT terminal coding agent. Vinci shows up inside Pi as a provider (we don’t ship or rebrand Pi), and a small character pack carries the Vinci voice into the CLI.

Every call routes through the Vinci gateway, so you get the character layer, free-tier metering, and ZDR — the same guarantees as the API and the web app.

1. Install the pack

The integration lives in getsimpledirect/vinci-pack.

git clone https://github.com/getsimpledirect/vinci-pack && cd vinci-pack
mkdir -p ~/.pi/agent/prompts ~/.pi/agent/extensions
cp AGENTS.md APPEND_SYSTEM.md ~/.pi/agent/
cp prompts/*.md ~/.pi/agent/prompts/
cp vinci-provider.ts ~/.pi/agent/extensions/   # the registered-provider path

Pin your Pi version — its model registry and config schema change per release. The pack is verified against @earendil-works/pi-coding-agent@0.79.8.

2. Get a key

In the Vinci app: Account → API keys → Issue a vinci_live_… key.

3. Configure & run

npm install -g @earendil-works/pi-coding-agent@0.79.8
export VINCI_API_KEY=vinci_live_...
 
pi --provider vinci --model vinci-piccolo

The provider points at https://vinci.getsimpledirect.com/api/v1 by default. To run against a local Vinci app (or self-hosted deployment), override the base URL:

VINCI_BASE_URL=http://localhost:3000/api/v1 VINCI_API_KEY=vinci_live_... \
  pi --provider vinci --model vinci-piccolo

Local / fully-owned lane

You can also run Vinci’s open weights entirely on your own machine (zero hosting cost, ZDR by construction) via Ollama or LM Studio — see the vinci-pack README .

Models

vinci-piccolo (light, recommended default for the CLI), vinci-bozza, vinci-tela, and ds-v4-flash. The footer shows real per-million-token cost — long, cache-stable sessions are where Vinci is cheapest.