CloudCode.ONE now supports Pi, the minimal coding agent harness
CloudCode.ONE now supports Pi
Pi (pi.dev) is a minimal terminal coding harness by Mario Zechner. The core ships four tools — read, write, edit, bash — plus a fast TUI, and everything else is something you add: TypeScript extensions, skills, prompt templates, packages. No MCP, no sub-agents, no plan mode unless you install them. It's the opposite design stance from the heavyweight CLIs, and it sends noticeably less context per turn as a result.
That makes it a natural fit for a pay-as-you-go gateway. You're not paying a subscription for a feature surface you don't use, and you're not paying for context you didn't need to send.
As of today, CloudCode.ONE ships an official Pi provider package.
Install
Two commands:
curl -fsSL https://pi.dev/install.sh | sh
pi install npm:cloudcode-pi-provider
The first installs Pi itself (Windows users: powershell -c "irm https://pi.dev/install.ps1 | iex"). The second installs our provider package into ~/.pi/agent/npm/, which registers CloudCode.ONE as a provider along with the current model catalog — no hand-written models.json, no base URL to copy around.
Add -l if you'd rather install the provider per-project instead of globally:
pi install -l npm:cloudcode-pi-provider
Authenticate
Grab a key from your CloudCode.ONE dashboard and export it:
export CLOUDCODE_API_KEY=cc-...
Put it in your shell profile so it survives a new terminal. Then start Pi:
pi
Cycle models with Ctrl+P, or pick one up front:
pi -m cloudcode/glm-5.2
What you get
The whole catalog, one key. Every model on the gateway shows up in Pi's model picker with correct context windows, token costs, and reasoning flags, so Pi's cost footer and compaction thresholds are accurate rather than guesses.
Automatic vision routing. This is the part that matters most for a harness like Pi. GLM-5.2 and DeepSeek V4 are text-only, so pasting a screenshot into a text-only session normally means either an error or a manual model switch. The gateway detects image content in the request and routes it to a vision-capable model for you — a single provider entry, no pi-vision-handoff-style extension needed, no vendor name hardcoded in your APPEND_SYSTEM.md.
Mid-session switching that actually costs what you'd expect. Start a refactor on a cheap model, hit Ctrl+P and move to a stronger one for the hard part, drop back down. Same key, same billing line, no second subscription.
Free tier for the boring turns. GLM-4.7-Flash is free on the gateway. Point Pi at it for commit messages, quick greps, and anything where you're mostly paying for tool loops rather than thinking.
Notes
Pi has no built-in permission system — by default it runs with the permissions of the user that launched it. That's a deliberate design choice, not an oversight, but it's worth knowing before you point it at a repo you care about. If you want boundaries, containerize it or install a permission-gate extension.
To update later:
pi update --extensions
If a model doesn't appear after install, check that CLOUDCODE_API_KEY is exported in the shell you launched pi from, then restart the session.