Feature request: Agent Skills parity in Claude Desktop chat / Cowork (Workflow tool, code-exec network, working files)
I maintain an Agent Skill built and tested in Claude Code that I would like non-developer colleagues to run in Claude Desktop chat and Cowork. The skill's pattern is generic and, I believe, common: collect data through a remote MCP server → analyze it with multi-agent orchestration → produce a report with bundled Python scripts.
Porting such a skill to Desktop chat / Cowork is blocked today by three gaps:
- No Workflow tool / multi-agent orchestration outside Claude Code. The skill relies on the Workflow tool (parallel
agent()batches, independent re-grading of outliers,budget.spent()for cost reporting). Desktop chat and Cowork have no equivalent primitive; falling back to sequential in-context analysis loses independence between agents and does not scale.
- Code-execution network access is not guaranteed (full/partial/none depending on user/admin settings). MCP servers that export large payloads out-of-band via single-use download URLs — a good pattern to keep bulk data out of the conversation context — cannot be consumed reliably from the code-execution environment.
- No persistent working files between steps. Multi-step script pipelines (collect → transform → aggregate → render) are awkward without a stable working directory across the skill's steps.
What already works well on Desktop: uploading the skill as a zip, and connecting the remote MCP server as a custom connector.
Request: bring Agent Skills in Claude Desktop chat / Cowork closer to Claude Code parity — ideally the Workflow tool (or an equivalent subagent fan-out primitive), plus predictable network access in the code-execution environment for MCP-issued download URLs. Even a documented supported subset would let developers ship skills built and tested in Claude Code to business users on Desktop/Cowork.