Reference

The two agentic clients Thinkube IDE carries, Thinkube Tandem Chat that brings Pi into the chat panel, the models and MCP servers they are set up with, what they keep inside the cluster, and how to add a Claude API key.

TL;DR

In the chat panel of Thinkube IDE, open the menu beside + and choose New Tandem (powered by Pi) Session, or run pi in a terminal. Ask: "which Thinkube services are running?". The client calls the list_services_minimal tool of Thinkube Control and answers from it, with a model served by the LLM Gateway. Nothing leaves the cluster. The agent is installed beside them and works with your Anthropic account. Read this page to see what each client is set up with, or to add a Claude API key.

The clients

Client Version License How you open it

Pi

1.0.0

MIT

pi in a terminal, the pi terminal profile, the task Pi - New Session, or New Tandem (powered by Pi) Session in the chat panel.

Claude Code

2.1.288

Anthropic’s terms

claude in a terminal, or the Claude Code panel. It signs in to your Anthropic account.

Each version is fixed in the Thinkube IDE image, and no client updates itself. The packages the clients depend on are listed in /opt/agent-clients/package-lock.json in Thinkube IDE; none is under GPL or AGPL. Pi also loads two extensions of its own package: permission-gate, which asks before rm -rf, sudo and chmod/chown 777 in a shell command, and subagent, which hands a task to a separate Pi with the scout, planner, reviewer or worker instructions. A subagent runs on the model of the Pi that starts it.

Thinkube Tandem Chat

Thinkube Tandem Chat is a Thinkube IDE extension that runs Pi as a chat of VS Code’s chat panel, with the same configuration as in a terminal. It needs no GitHub account.

  • A chat is a session. Each chat keeps its agent session: the next message continues it, also after a reload, and the chat lists in Sessions.

  • Questions in the chat. A confirmation or a permission request of the agent is asked in the chat, with its choices.

  • Edits you can undo. A file Pi changes shows as a change with Keep and Undo. Pi waits before each edit until VS Code has recorded the file.

  • A model per chat. The chat panel’s model picker lists the models Pi can call. A new chat starts with the agent default; a model you pick stays with that chat.

Models

Pi uses the LLM Gateway at https://llm.<your domain>/v1. It offers the models of the model catalog marked agent that the gateway serves, and the catalog’s agent default even when it is not loaded. Each is named by its gateway alias.

Alias Model

thinkube-fast (the agent default)

Qwen 3.6 35B-A3B, NVFP4 (nvidia/Qwen3.6-35B-A3B-NVFP4)

thinkube-deep

Qwen 3.8 27B, NVFP4 (nvidia/Qwen3.8-27B-NVFP4)

thinkube-flash-next

Qwen 3.8 Flash Next on Flare, for DGX Spark (local-inference-lab/Qwen3.8-Flash-Next-FLARE)

A model answers once it is loaded: ask your agent "load nvidia/Qwen3.6-35B-A3B-NVFP4", or load it on the LLM Gateway page of Thinkube Control. Pi reads the list from the gateway when it starts, with the context each model is served with, so a model loaded later appears in a new chat. A call to the agent default while it is not loaded answers with what to load.

The aliases, the agent mark and the agent default are fields of the model catalog: aliases, capabilities and default_agent in models.json. Your own metadata repository can add or change them; its claim to an alias or to the default wins over the platform’s. Pi signs in to the gateway with the Thinkube Control API token that Thinkube IDE carries as THINKUBE_API_TOKEN.

MCP servers

Every client is connected to the same servers.

Server What it offers

thinkube-control

Thinkube Control: services, apps, templates, models, images, secrets and notebooks

tk-package-version

The latest versions of npm, PyPI, Go and Cargo packages and container images

playwright

A headless browser to open and test web apps

Tools that change the platform

A tool that only reads, such as list_services_minimal or get_llm_models, runs at once. Every other tool of these servers, such as delete_model, uninstall_optional_component, toggle_service or restart_service, asks you first.

Client How it asks

Pi

A confirmation with the tool’s arguments, in the terminal or in the chat. pi -p and pi --mode json have nobody to ask, and the call is refused.

Claude Code

Its own permission settings. Thinkube IDE allows every tool of thinkube-control and tk-package-version without asking.

What stays in the cluster

The clients send your prompts and files to the LLM Gateway only. Each other connection they could make is switched off.

Client Connection Switched off by

Pi

Install report to pi.dev

PI_TELEMETRY=0, "enableInstallTelemetry": false

Pi

Version check and model catalog refresh from pi.dev

PI_OFFLINE=1

Claude Code

Telemetry and error reports

DISABLE_TELEMETRY, DISABLE_ERROR_REPORTING

Claude Code

Self-update

DISABLE_AUTOUPDATER

Claude Code sends your prompts to Anthropic: it is the client for work with Claude.

Instructions

Each client reads the platform’s operating instructions from one global AGENTS.md: the git, infrastructure, comment, CI/CD and operations rules, among others.

Client File

Pi

~/.pi/agent/AGENTS.md

Claude Code

~/.claude/AGENTS.md, imported by ~/.claude/CLAUDE.md

Thinkube writes these files each time it configures Thinkube IDE. Put your own instructions in a project’s AGENTS.md, which every client reads too.

Add a Claude API key

Pi uses Claude with an Anthropic API key, created in the Anthropic Console. Run /login, choose Anthropic, then the API key option, and paste the key. Choose a Claude model with /model. Pi also reads the key from ANTHROPIC_API_KEY when it is set in the terminal it starts in.

Limits

  • An edit to an existing file outside the workspace folders shows no change, because VS Code does not watch that file.