Skip to main content
Read sandbox MCP resources to retrieve the bead skill and complete output from large retained tool results without repeating operations.

​
Read the skill

Use resources/list and resources/read to load:
The bead prompt is also available through prompts/list and prompts/get. Client support determines whether prompts appear as slash commands. Reading a resource does not install a native filesystem skill. Clients that negotiate the skills extension can use skills/list, skills/get and resources/directory/read. Directories are skill://bead and skill://bead/references. Manifests include YAML frontmatter, SHA-256 digests and raw byte sizes. The same files are available through bead skills get bead and bead.run/SKILL.md.

​
Continue large output

Tool output targets 64 KiB by default. page_bytes accepts 1–1040384 bytes, accounting for JSON escaping within the wire limit. Inspect total_bytes and retained cursor fields.
This is a tools/call parameter object. Replace the cursor with the one returned by your operation. A retained result cursor is distinct from the upstream tool-list pagination cursor. Oversized JSON is returned as json-fragment pages. Concatenate the fragments before parsing, or download the complete artifact. Following a cursor does not rerun the original operation.

​
Downloads and expiry

Native stdout, stderr and file bytes remain retrievable for 600 seconds. Download URIs require the original caller’s current bearer credential. A handle alone grants no access. Guest transport credentials are withheld from tool output. Retrieve or save required output before expiry. See command recovery for uncertain execution and worktree recovery for worktree-specific retention.