All Templates
Pi Coding Agent logo

Pi Coding Agent

AI Agent

Coding agent CLI with session history and a browser terminal

Deploy Now

README

Coding agent CLI with session history and a browser terminal.

Overview

This template runs the Pi coding agent inside a container that exposes a browser terminal. You open a URL, authenticate, and get a bash shell with the pi CLI already installed. Upstream describes it as a coding-agent CLI with read, bash, edit and write tools plus session management: so a session you start now can be picked up later from the same URL.

The image is built from the Dockerfile in this directory: node:24-bookworm-slim (pinned by digest) plus ttyd 1.7.7 (verified against a pinned SHA-256) and @earendil-works/pi-coding-agent pinned to an exact version. The base image, ttyd and the agent CLI are fixed, so the CLI you get is the one the template version names. The rest is not: every Debian package is installed by name from Debian's live bookworm repositories, gh from upstream's rolling apt repository and bun from npm, so all of them can move between two builds of the same template version. In front of ttyd sits the InstaCloud sign-in page, insta-gate from this repository's gate/, also verified against a pinned SHA-256. ttyd listens only inside the container and never receives your credentials. Template versions before 0.8.3 logged the credential on every start, and upgrading does not remove those lines from the log history: if you ran one, set a new ADMIN_PASSWORD before you upgrade.

What you get by hosting it

  • An HTTPS URL for the terminal, with no port forwarding or tunnel to manage.
  • A persistent volume mounted at /data. HOME is set to /data/home, so session history, CLI config and any repositories you clone survive restarts, redeploys and version upgrades.
  • The terminal credentials kept as service variables rather than baked into the image, so you can change them later without rebuilding anything. They are yours, not ours: the template ships no credential of its own, and both values are visible in the deploy form and in the service's variables.
  • Deploys are health-gated: a container that does not answer is rolled back to the last healthy image instead of leaving you with a dead URL.
  • A terminal that outlives the tab. Every tab attaches to the same tmux session, main, so closing the tab or dropping the network does not stop an agent mid-task, and opening the URL again picks up where it was. The machine itself still scales to zero once nothing is connected, which ends the session. To keep agents working with nobody connected, run insta compute always-on on <service>, which bills for the uptime.
  • Common tools preinstalled: gh, git, curl, ripgrep, jq, ssh, rsync, unzip, zip, less, tree, fzf, htop, nano and vi, python3 with venv and pip, bun, and the C and C++ toolchain that npm install needs to build a package with a native addon. For working out what is wrong from inside the box: lsof, ss, dig, netstat, socat. A machine that mounts the platform toolbox at /.insta/tools/bin has a second gh there, and a git credential helper already pointing at that one keeps working.

What you need before deploying

  • A username and a password of your choosing for the terminal sign-in. There is no default: the deploy form starts with both fields empty and will not submit until you fill them.
  • A model provider key if you want it present from the first boot. Any one of the three below is enough, and none is also fine: you can add a key from inside the terminal instead.

Configuration

VariableRequiredWhat it does
ADMIN_USERNAMEyesUsername for the InstaCloud sign-in page in front of the terminal. You choose it.
ADMIN_PASSWORDyesPassword for the InstaCloud sign-in page in front of the terminal. You choose it.
ANTHROPIC_API_KEYnoAnthropic key, for Claude models.
OPENAI_API_KEYnoOpenAI key, for GPT models.
OPENROUTER_API_KEYnoOpenRouter key, for whichever model you route to.

Both credentials are required and neither has a default, so the deploy form starts empty and refuses to submit until you supply them.

Pi reads a provider key straight from the environment and picks its model accordingly. It recognises more than thirty providers (Gemini, DeepSeek, Groq, Mistral, Kimi and so on); the three above are the ones this template puts on the deploy form, and the rest work just as well as service variables or configured from inside the terminal.

Set by the template, not by you: HOME=/data/home (puts your home directory on the volume).

Pick the password like it guards a shell, because it does. What it protects is a root shell that can run anything and holds whatever API keys you gave it, so whoever has the URL and this password has all of that. Both fields can be changed later from the service's variables.

After deploy

  1. Open the service URL. An InstaCloud sign-in page asks for the ADMIN_USERNAME and ADMIN_PASSWORD you deployed with. The session lasts 30 days, and changing either variable signs every browser out.
  2. You land in a bash shell in /data/home.
  3. Run pi. Configure a model provider key if you did not set one as a variable.
  4. Configuration and session history persist. Because HOME is on the volume, ~ survives restarts, so a session started before a redeploy is still there afterwards.
  5. Clone your repository into /data/home (or anywhere under /data) so your work persists too. Files written outside /data are lost when the container is replaced.

Links

Services & Specs

pi
Web service
Image
ghcr.io/insforge/insta-oss/templates/pi:0.11.0
Port
7681
Healthcheck
/

Variables

You supply 2 variables before the first deploy.

Required

ADMIN_USERNAME

Username for the browser terminal sign-in

ADMIN_PASSWORD

Password for the browser terminal sign-in

Optional (3)
ANTHROPIC_API_KEY

Anthropic API key, for Claude models

OPENAI_API_KEY

OpenAI API key, for GPT models

OPENROUTER_API_KEY

OpenRouter API key, for whichever model you route to