Skip to main content
The agent workspace is a private Linux sandbox the agent can use while it works on your request. It can install packages, write and run Python or Node scripts, keep files between steps, and publish the results back to your canvas. Use it for work that doesn’t fit in a single Code node: analysing a spreadsheet with pandas, generating geometry with trimesh, scraping and cleaning data, or building a project with several files.
The workspace is experimental and needs a Hobby plan or higher. It builds on agent teams, so turn that on first.

​
Turning it on

  1. Open the account menu and choose Preferences.
  2. Under Features, turn on Agent teams (experimental).
  3. Turn on Agent workspace (experimental).
The switches take effect right away. Open any saved workflow and chat as usual. The agent decides when a task needs the workspace, or you can ask for it: “use the workspace to analyse this CSV”.

​
What the agent can do

Inside the workspace the agent can:
  • Run commands: any shell command, including installing packages with pip, uv or npm
  • Write and edit files: create files, make targeted edits, or apply a patch across several files at once
  • Read and search: list folders, read files and search their contents
  • Look at images: view a plot or render it made, to check the result before sharing it
  • Publish to the canvas: send a finished file to your canvas, where it appears as a node
The sandbox has internet access, so the agent can download data and install packages.

​
What’s installed

The workspace comes with Python 3.12, Node, uv, git and ripgrep, plus these Python packages: Anything else the agent installs lasts until the workspace goes to sleep.

​
Getting results onto the canvas

You can’t see inside the workspace, so the agent publishes what it makes. A published file (up to 8 MB) is stored in your Runchat account, and the agent adds it to the canvas as an Input node:
  • Images and video show as media
  • 3D models (.glb and .obj) open in the model viewer. Materials and textures from an OBJ’s .mtl file are not loaded.
  • Data and documents such as CSV, JSON, PDF and HTML are added as files you can connect to other nodes
If the agent finishes without publishing, ask it to “publish the result to the canvas”.

​
How long it lasts

Each chat gets its own workspace. It starts the first time the agent uses it and goes to sleep five minutes after its last use.
Files in the workspace are deleted when it goes to sleep. Ask the agent to publish anything you want to keep.
When the agent is working as a team, the coder and modeller teammates share the lead agent’s workspace, so they can work on the same files.

​
Cost

The workspace is billed by the minute while it is awake, at about 1 credit per minute. That includes the five idle minutes before it goes to sleep, so a short task costs around 6 credits on top of the agent’s own usage. If you run out of credits the workspace stops. See Credits.

​
Limits

  • 1 GB of memory, a quarter of a CPU core and 4 GB of disk, so very large datasets or meshes may not fit
  • No GPU
  • Files up to 8 MB can be published; images the agent views must be PNG, JPEG, WebP or GIF up to 5 MB
  • The workspace is only available in chats on a saved workflow, not from the Agent node, MCP or the API

​
Next steps