A loadout is a ready-to-run job for an AI agent: the playbook, tools, guardrails, examples, and success criteria packaged into one portable markdown file. MCP gives agents tools — a loadout tells them how to use those tools to win the job. One file. Any MCP client.
Workflows do steps. Loadouts do jobs.
The market has settled the easy question — AI can control your apps through MCP. The hard question is whether it knows how to run the job well: a workflow says "when X, do Y"; a loadout says "here's how to win this whole outcome."
No context, no gate. "Done" just means it ran out of steps — it fires again tomorrow whether it worked or not.
Interviews you first, stops at the gate for approval, and proves the job is done with a verifiable check.
Everything else in the format bends. These don't.
The file opens with a config interview — and the agent only asks what it can't discover from the file, the conversation, its memory, or the account itself. In a chat, these are questions; in a dashboard, the same lines render as a form. One artifact, two front-ends.
Before anything irreversible — a send, a post, an automation — the workflow stops and a human approves. A loadout without a gate is a script, not a loadout.
“Done means” must be checkable — name the read tool that proves the work happened. No vibes-based completion.
Seven parts, in order. Full section reference in the spec README.
# <Name> — Loadout ```yaml loadout: <slug> # required — stable id mcp: <server url> # required — the MCP this file drives tier: <access tier> # optional cadence: weekly # optional — rhythm if run as a recurring loop ``` **The job:** one paragraph. The outcome, in the operator's language. ## Before you run — ask the user Only ask what you don't already know. Check the Config, the conversation, memory, and the account first. Confirm inferences in one line. 1. **Who is this for?** (hint) 2. **What's the asset?** (name + URL) ## Config ```yaml audience: # keys become stable form-field ids asset: ``` ## Steps ### 1. See — call `a_real_tool`, study the signal ### 2. Decide — rank, filter, dedupe against history ### 3. STOP — approval gate ⛔ Present everything. Nothing irreversible happens past here without a human. ### 4. Act — only what was approved · ### 5. Close the loop for next run ## Never - The hard boundaries. No fabrication. No acting without approval. ## Done means - Verifiable — name the read tool that proves it (`list_actions`, …)
Not a hypothetical — this file runs against a production MCP. It interviews for config, studies live market signal, learns the account's voice from its own post history, drafts with the capture automation attached, and stops at the gate for approval.
Connect the MCP named in the file's header, then paste the whole file into your agent with this:
Works in Claude, ChatGPT, Cursor, or any MCP client. Dashboard runtimes render the same file as a form + checklist + approval queue — Apex by LeadShark is the reference runtime the format was proven against.
Paste a loadout.md — checked against the v0.1 spec, entirely in your browser. Nothing leaves this page.