Kiro CLI 2.27 MCP prompts as slash commands: what Sume exposes

Kiro CLI 2.27.0 (2026-10-01) exposes workspace, global, and MCP prompts as slash commands with argument hints in CLI V3 (read 2026-10-03), but Sume's hosted MCP documentation describes tools, not prompts. Do not expect typing / in Kiro to list Sume commands because the server is connected. What you can do is write a workspace or global saved prompt that tells Kiro how to call Sume's tools, and then it appears as a slash command through the same feature.

The Kiro entry is on the changelog. The Sume tool inventory is in MCP tools and gates. If an MCP prompt feature ships on Sume's server later, its docs will say so; absent that, plan for tools.

What 2.27 added

The release also adds a setting for how the main chat delegates, directly to sub-agents or through Workflows, and #[[file:...]] references with line selectors and #[[folder:...]] references in live steering. Saved prompts as commands is the piece that matters for repeat Sume work, because it lets a team turn a good generation procedure into a command everyone runs the same way.

A saved prompt that uses Sume tools safely

The prompt below is the kind of thing to save at workspace level. It names real hosted tools, keeps the first step read-only, and makes the paid step conditional on a preview. Arguments depend on the tool, so the prompt tells the agent to read tools_schema rather than guessing field names.

Make one image for: {{brief}}

1. Call Sume tools_schema for generate_image and read the fields.
2. Call generate_image with dry_run=true and show me the estimate.
3. Wait for my yes. Then call it again without dry_run, with a new
   idempotency_key and max_spend_usd set to the estimate rounded up.
4. Call jobs_wait on the returned job id. If wait_slice_expired,
   call jobs_wait again. Do not call generate_image a second time.
5. Report the media.sume.com URL and the job id.

Prompts, skills, or tools

Three mechanisms carry instructions, and Sume touches two of them. Tools are the executable surface of the hosted server. A saved prompt is text your team owns, and it is the right place for house rules like a spend ceiling. Sume's CLI also ships a packaged skill: sume skills install writes into .agents/skills or .claude/skills, and sume skills export sume lets you read the source first. That route is for agents that run shell commands.

If your Kiro setup reads skills from one of those folders, the skill route may fit; if not, a saved prompt is the portable choice. Either way, keep keys out of the prompt text, and prefer the OAuth connector or an environment variable for credentials.

Naming and sharing the command

A saved prompt is only useful if people can find it, so name it for the job rather than the tool, such as a product shot or a caption pass, and put a one-line description at the top. Keep the argument list short; Kiro shows argument hints, and a prompt that needs seven arguments will be skipped for ad hoc typing.

Workspace prompts live with the project, which makes them reviewable. A change to the spend ceiling in a prompt is a text diff someone can read in a pull request, which is more than you can say for a setting buried in a client. Global prompts live with the person, so use them for personal habits like always previewing first. When the two disagree, the narrower scope usually wins in tools like this, but check Kiro's own documentation for precedence before relying on it.

Keep one rule non-negotiable in both: the paid step needs a fresh idempotency_key every time a new job is wanted, and the same key when retrying one. Reusing a key across different jobs would dedupe a genuinely new request into the old one.

Testing the command before sharing it

Run the saved prompt once with a cheap call and read the transcript. You want to see the schema read first, the preview second, an approval pause third, and one create fourth. If the agent skips the preview, tighten the wording; if it creates twice, check that the retry instruction says to wait again, not create again.

Then try it with a deliberately bad brief, such as an empty one, and confirm the agent stops. A prompt that handles the failure path is the one worth putting in front of teammates.