pi-basert

BaseRT pi extension. Auto-discovers models from a running basert serve server and registers them as the basert provider in pi.

pi install git:github.com/basecompute/pi-basert

Install

From the shell:

pi install git:github.com/basecompute/pi-basert

This clones to ~/.pi/agent/packages/pi-basert/ and adds an entry to your pi settings. Every future pi invocation auto-loads it.

From inside an interactive pi session:

!pi install git:github.com/basecompute/pi-basert

Then run /reload (or restart pi) to load the extension.

Dev mode:

git clone https://github.com/basecompute/pi-basert ~/code/pi-basert
pi -e ~/code/pi-basert/index.ts

-e loads the extension only for the current session, useful while developing.

Environment Variables

This extension supports the following environment variables:

• BASERT_BASE_URL (Default: http://localhost:8080/v1)
• BASERT_API_KEY (Default: no-key) — sent as Authorization: Bearer … to both the OpenAI chat endpoint and the discovery endpoints (/v1/models, /props). Required if you started the server with --api-key.

Usage

# 1. Install BaseRT
curl -LsSf https://basecompute.co/install.sh | sh

# 2. Pull a model and start the server (serves on port 8080 by default)
basert pull Qwen/Qwen3-0.6B
basert serve Qwen/Qwen3-0.6B

# Optional: point at a remote BaseRT server instead of local
export BASERT_BASE_URL="https://basert.example.com/v1"
# Optional: if the server was started with --api-key
export BASERT_API_KEY="sk-…"

# 3. Launch pi in another terminal
pi

# 4. Inside pi - search "basert" to browse your loaded models
/model

In the picker, multimodal models are tagged (image) and not-yet-resident models (from basert serve --model-dir) are tagged (unloaded). Selecting an (unloaded) model loads it into VRAM on demand (you'll see a loading… status), and pi sizes its context budget to each model's reported window.

Commands

• /basert-version — prints the running BaseRT server's build version and target.

How it works

• GET /v1/models lists the models currently loaded by basert serve; each is registered under the basert provider via pi's openai-completions API. Each entry carries meta.n_ctx (context window) and architecture.input_modalities (e.g. ["text","image"]), so pi sizes context and flags image-capable models without a second round-trip.
• GET /props?model=<id>&autoload=1, fetched the first time you select a model, loads it if needed and supplies the raw chat_template (used to detect an enable_thinking reasoning toggle) plus default_generation_settings.max_context / max_tokens. The loaded flag on /v1/models drives the (unloaded) tag.

When a model exposes an enable_thinking template toggle, pi's thinking level is sent as chat_template_kwargs.enable_thinking, which basert serve honors.

Troubleshooting

Error: 404 Model not found — pi is requesting a model your basert serve instance isn't currently serving (usually a model selected in a previous session, e.g. an old *.base file). The provider only exposes what the server actually serves, so run /model and pick a model from the list. Confirm what's available with curl http://localhost:8080/v1/models.