Skip to content

Claude Code and Posit Assistant

This page describes the coding agents in the IDE sessions, what the workspace gives them, and how a customer adds their own instructions. Every IDE session has Claude Code in the terminal, and RStudio also has Posit Assistant in its pane. Both get the same knowledge from the workspace, so you can ask either one to start a project, deploy it, read a failed build or set up a schedule, and it knows the commands and conventions of this workspace.

Accounts

The workspace ships no AI credentials. Claude Code asks each user to sign in with their own Anthropic account on first use: run claude in the terminal, sign in on the page it opens in the browser, and paste the code shown there into the terminal. Posit Assistant asks in its pane for a Posit AI login or an API key of a model provider, and keeps it in the user's home. By default, Claude Code accepts any Anthropic account, a private one included. To accept only the accounts of your organisation, set claude_login_method and claude_login_org in velonde.env. If a customer wants a central API key or Bedrock for Claude Code instead, it goes into the workspace's configuration, and we set this up with them.

Where your code and data go

A coding agent sends what it works with to the model behind it: your prompts, the files it reads, the output of the commands it runs, and, through btw, objects from your R session. With the default setup, that goes to Anthropic for Claude Code, and to Posit or the model provider whose key you entered for Posit Assistant. It is handled under the terms of the account the user signed in with. Nothing goes to us, and deployed apps do not run through the agents.

How much leaves your network is a choice of the agent and the model:

  • A company account or a central key. For company code, let users sign in with your organisation's Anthropic account, or have Claude Code use one central API key or Amazon Bedrock in a region you choose. We set up the central options with you.
  • A different agent or a model you host. The workspace is not tied to Claude Code and Posit Assistant. On request we set up another coding agent, such as OpenCode, in the IDE images. With an agent that talks to a model you host yourself, nothing leaves your network.
  • No AI. With ai=false the workspace has no coding agents at all, as described below.

What the agents get from the workspace

Managed instructions. Every session loads a short text before the user's own CLAUDE.md. It says where files persist, what the deploy command is and how projects are laid out. The parts about R, Python and deploying are included only on a workspace whose features setting includes them.

Rules and skills. Rules apply when the agent works on matching files. They cover the R conventions (code style, formatting with Air, how imports are declared), the Python conventions (uv for packages, code style), and tell Claude Code to use the running R session through btw. Skills are loaded when a task needs them. The velonde skill is the full reference for velonde deploy, velonde.yaml, the kinds, pipelines and APIs, and for reading and fixing a build log. The project-claude-md skill describes how to write a project's own CLAUDE.md. The rules and skills are part of the workspace's configuration. At every session start, the workspace unpacks them and links them, read-only, into the user's Claude Code configuration, and into Posit Assistant's skills directory and its AGENTS.md.

Plugins. Posit's skills for Shiny, the r-lib packages and Quarto are installed into the user's Claude Code at session start, the Quarto skills also on a workspace with Python only. They come from a marketplace built into the image, so a session start does not need network access to install them. When an image update brings a new version of the marketplace, the next session start updates the plugins. A plugin that the workspace drops from its list, for example because a feature was turned off, is uninstalled. A plugin that the user installed themselves stays installed.

Permissions. Claude Code starts in its auto mode. In that mode Claude Code's own safety check decides about a command that neither of the workspace's two lists names. The agent runs the commands of the first list without asking: velonde deploy, velonde rollback, velonde init, velonde config, velonde sysreqs, the workspace's commands that only read (velonde apps, velonde history, velonde logs, velonde runs, velonde env list, velonde git status, velonde mcp), air format, uv lock, uv sync and uv add. It always asks before the commands of the second list: velonde delete, velonde run, velonde apikey, velonde env set and velonde env unset, velonde mcp add, velonde new, velonde clone, velonde owner, velonde packages, velonde git create, bind, unbind and logout, velonde ide-packages add and remove, and velonde config set or unset for audience, access-groups, access-users, access-clients, maintainers and maintainer-groups. These commands show or change secrets, cannot be undone, or change who can reach an app. The lists name commands only. The agent can also edit velonde.yaml and deploy. A deploy that makes an app external or public, or that removes its access lists, needs --confirm-access, and the agent always asks before it runs a deploy with that option. For every other access setting, read what the agent changed in velonde.yaml before it deploys. The default model is the newest Opus. These are managed settings that a user cannot override. Posit Assistant gets the same commands in its own settings at every session start. A rule that you add there yourself stays, and a rule that the workspace drops from its list is removed.

MCP servers. In RStudio and in VS Code, Claude Code talks to the user's R session through btw: RStudio's console, or an R terminal in VS Code. It can inspect objects, run code in the session and read help pages. The r-btw server also appears in JupyterLab, because the IDEs share the home directory, but it has no tools there. In every IDE, the workspace registers each deployed MCP server that the user's groups have access to, under the server's name, with the session as the identity, and the servers your admins list for everyone (MCP servers for everyone). In RStudio, Posit Assistant gets the same servers. A server of your own with the same name stays as it is.

Your own CLAUDE.md. The workspace creates it once, with a short note on working style. After that the file is yours to edit.

A setup step at session start does all of this. It never blocks the session, and it logs what it did, or could not do, in ~/.claude/velonde-setup.log.

Adding your own instructions

A customer's additions go into the claude-code/ directory of the overlay:

File What it contains Takes effect
CLAUDE.md Text appended to the managed instructions of every session: your systems and data, naming conventions, what the agent must never do, whom to ask. Keep it short, because it uses context on every turn. at the next apply
plugins.txt Additional marketplaces and plugins, in the platform's format. at the next apply
rules/, skills/ Your rules and skills, linked into every session alongside the platform's. Together with the platform's they may take up to 700 KB packed. at the next apply, in new sessions
claude-marketplaces.txt next to velonde.env Marketplaces to build into the images, so plugins install without network access. at the next image build

Each project template in the overlay carries its own project CLAUDE.md. Conventions for one kind of project belong there rather than in the workspace-wide file.

A workspace without AI

With ai=false in velonde.env, the workspace builds the IDE images without Claude Code, Posit Assistant and the R packages that drive them, and ships none of the content above. The tiles' descriptions change accordingly. Everything else about the workspace stays the same.