Run agents
Prompt Studio runs coding agents that you install, such as Claude Code, Codex, and OpenCode. Install an agent, then start a session.
Install an agent
Each new project has a harness extension for each supported agent. A harness connects Prompt Studio to the agent’s own command-line tool. It does not install the agent or sign you in.
| Agent | Command Prompt Studio looks for | Supported versions | Harness docs |
|---|---|---|---|
| Claude Code | claude |
2.1.203 and newer | Claude Code |
| Codex | codex |
0.157.0 and newer | Codex |
| OpenCode | opencode |
2.2.5 and newer | OpenCode |
An older version shows as unavailable with the reason, such as the version it found and the version it needs. Update the agent, then restart the runtime.
Install the agent and sign in as its own documentation describes. Then check what Prompt Studio finds:
pst agents list
The list shows each harness, its full ID, and whether its command is installed.
Start a session
A session is one conversation with an agent. In the dashboard, choose New conversation on the project’s Start page. Pick a harness and a model, type your request, and send it.
If no enabled harness is available, a single No agent available alert replaces the message field and its controls. The workspace container and header stay visible. Choose Browse extensions in the alert to open the project’s Extensions page. An enabled harness extension that is unavailable shows the warning indicator in its list row and detail page. Open that indicator to read the reason supplied by the extension. The composer returns after the next successful availability check. Your draft text and attachments stay saved while it is hidden.
In an existing session, pending questions and confirmations stay available beside the warning. Answers you have already entered stay in the form when availability changes.

The conversation opens in a side panel so you can keep a tool visible while talking to the agent. The image shows a draft request before sending it. It asks for a small tool and describes what the person should be able to do with it.
Before sending:
- Check the workspace selector above the message field. Project folder uses the project’s own files; choose another ready workspace if you want the work elsewhere.
- Use the model selector below the message field to choose the agent and model available on your machine. The pictured model is an example, not a requirement.
- Describe the result you want and how you will check it. For a new tool, follow Ask an agent to build your tool, including skill setup.
From the terminal, run this inside the project folder:
pst sessions create --prompt "Summarize this folder" --agent pstdio.harness-claude-code.harness.claude-code
--agent takes the full harness ID that pst agents list shows. Leave it out to use the project’s default.
Follow the session and continue it:
pst sessions list
pst sessions stream --id <session-id>
pst sessions follow-up --id <session-id> --prompt "Now add a README"
pst sessions stop --id <session-id>
When too many sessions run at once, new ones wait with the status queued and start in order. Set the limit in Settings → Runtime → Max concurrent sessions. Leave it blank for no limit.

Choose Save after changing the concurrency limit. This limit applies across projects. Default model sets the starting choice for new sessions in the current project; you can choose a different model in a session’s composer.
Answer approval requests
Some agents ask before they run a tool or change a file. The session shows the request, and you approve or deny it there. From the terminal, pass the session ID and the request ID:
pst sessions approve --id <session-id> --approval-id <approval-id>
pst sessions deny --id <session-id> --approval-id <approval-id>
Choose where the agent works
Sessions run in the project folder by default, and they share its files. In a Git repository, you can give a session its own branch and folder instead. This is called a worktree workspace:
pst workspaces create --provider pstdio.worktree
# Created workspace WS-1 at ...
pst sessions create --prompt "Try the refactor" --workspace-id WS-1
pst workspaces create prints the new workspace’s short name, such as WS-1. --workspace-id accepts that short name or the workspace ID that pst workspaces list shows. The repository needs at least one commit.
Read Projects and workspaces to choose between them.
Give the agent skills
Extensions can ship skills: instructions that teach an agent how to use a tool. Run this inside the project folder to install the skills of your enabled extensions for one agent:
pst agents setup claude-code
This copies the skills into the project folder, where the agent reads them. Add --global-skills to install them in the agent’s global settings folder instead. Run it again after you turn on a new extension. For example, Planner adds skills for its ticket workflow, and Prompt Studio Skills adds a skill for writing extensions. Prompt Studio also copies the skills into each workspace it prepares for a session.
The installed skill folders are copies. To change a skill, edit its source in the extension that ships it, then reinstall or reload that extension. Edits to an installed copy can be lost:
- From the release after 0.43.x,
pst agents setupinstalls only missing skills and keeps existing skill folders. The automatic copy updates only files that Prompt Studio wrote and that nobody edited. To replace the installed folders of enabled skills with their source, and discard local edits, runpst agents install-skills claude-code --force. - In 0.43.x,
pst agents setupand the automatic copy replace installed skill files with the source every time, so local edits in them are lost.
To learn what the host and the harness each do, read Agents and harnesses. For every session command, see CLI sessions.
If an agent does not start, see Troubleshooting.
Updated at: 2026-11-05