Scout

AI agents & MCP

Scout is built to be driven by the agent that wrote the feature: it scaffolds onboarding files for coding agents, runs a co-author loop, and exposes an MCP server — with a verification-integrity rule that forbids claiming a pass without a real run.

Scout’s main flow is meant to be driven by the agent that built the feature: the agent writes the scenario in natural language (never the script — the script is born from the verified execution) and triggers verification. This is the loop closing on itself — the same agent that produces the code is held to the outcome it promised.

What scout init scaffolds

On scout init, Scout writes onboarding artifacts so any coding agent in the repo can help:

  • AGENTS.md (repo root) — the single source of truth for how an agent uses Scout. Written inside a managed block (<!-- scout:start -->…<!-- scout:end -->), so it refreshes on re-init without clobbering anything else in the file.
  • .claude/skills/scout/SKILL.md — a Claude Code skill, scoped to Scout/QA intent; points at AGENTS.md.
  • .cursor/rules/scout.mdc — a Cursor rule, glob-scoped to **/*.scout.md + scout.config.json; points at AGENTS.md.

Re-running init refreshes all three — init is the upgrade path. They’re scoped, not always-on, so they stay out of the way in repos where Scout is incidental.

The co-author loop

  1. You describe a flow you want covered, in plain language.
  2. The agent writes the .scout.md directly (richer than the CLI scaffold — it has full repo context: routes, components, copy, auth).
  3. The agent verifies with scout go.
  4. The agent reports the real verdict — verbatim. It never declares a scenario working without running it.
  5. You iterate together until it’s verified, then commit the .scout.md and recorded script.

Verification-integrity rule (baked into AGENTS.md): an agent must never claim a scenario passes without running scout go and reporting the actual verdict. A verdict is trustworthy only because it came from a real run.

MCP server

Expose Scout to a coding agent (Claude Code, cloud sessions) over MCP:

// .mcp.json of the target project
{
  "mcpServers": {
    "scout": { "command": "npx", "args": ["scout", "mcp"] }
  }
}

Exposed tools: scout_list_scenarios, scout_create_scenario, scout_run, scout_report, scout_get_run_report. The scout_run tool takes a baseUrl param (same precedence as the CLI flag).