Prompt access through MCP
Let an agent use your prompt library through Subshell MCP.
Before you start
Run an agent with Subshell MCP available. Follow MCP configuration for setup. The pane token identifies the subshell's owner; prompt access is based on that account.
Prompt tools
| Tool | Behavior |
|---|---|
list_prompts | Lists the owner's prompts and other users' shared prompts with descriptions and short previews. |
get_prompt | Reads a visible prompt's full body by ID. |
create_prompt | Saves a prompt belonging to the pane owner; sharing defaults to false. |
update_prompt | Changes the owner's description, body, or sharing flag. |
delete_prompt | Deletes the owner's prompt permanently. |
Use IDs returned by list_prompts, then call get_prompt to read the complete instructions. A preview is not the full task. Reading a prompt does not automatically submit it to another agent.
There are no dedicated prompt-stack MCP tools. The prompt tools above operate on individual saved prompts; do not treat a stack ID as a prompt ID. For authorized API automation, stack endpoints are under /api/prompts/stacks; see Build an API integration for authentication and schema discovery.
Ask an agent to save instructions
For example, tell the agent:
Save our repository review instructions as a private reusable prompt named
"Review changes". Do not include credentials or share it with other accounts.The agent can create and edit your entries and read other accounts' shared entries. It cannot edit another owner's prompt. A shared: true write publishes the text to every account on the instance; agents should leave it private unless sharing was requested.
Deleting through MCP has the same consequences as deleting in the dashboard, including removing the prompt from stacks and providing no undo.
Launch helpers with task text
An agent can read a prompt and supply its body when creating a helper using an eligible preset. It should choose whether to append to or replace the preset's saved task, inspect promptDelivered, and clean up the helper when done.
See Use and manage presets, Coordinate agents, and MCP tool reference for the launch and cleanup contracts.
Edit on GitHubLast updated on