Skip to content
AssureSwarm Docs

Agent Operating Guide

/coach-setup

One-time per-workspace setup that detects your Canvas MCP server, confirms your tenant, writes the config every other Audit plugin skill reads, and offers to schedule the autopilot.

Bootstraps a workspace so the plugin knows which tenant to operate on. The skill detects the AssureSwarm MCP server available in your Claude environment, confirms which company instance is yours, probes it with get_schema and get_current_context to confirm it’s reachable, then writes .coworkcanvas/config.json plus a schema snapshot. Every other skill in the plugin reads that config. Once connected, it also offers to schedule /coach-autopilot: the recurring pass that drafts your ready steps and requests missing inputs. It’s local workspace setup: nothing is written to your tenant, so there’s no suggest_change step. Re-running is safe: it refreshes the snapshot while preserving your redaction policy, installed-artifact records, and autopilot schedule.

When to use

Section titled “When to use”

The first thing you run in a new workspace, before any other skill: it records the tenant every skill needs. Re-run it after a schema change to refresh the snapshot, or when you switch to a different tenant.

Example

Section titled “Example”

/coach-setup scans your environment for a AssureSwarm MCP server, confirms the subdomain (asking if it can’t infer it from context), probes get_schema and get_current_context, snapshots the schema to .coworkcanvas/schema.json, and writes the config: printing one line like Canvas connected: acme.assureswarm.com (11 item types in schema). If the tenant has no item types yet, it offers to seed a /coach-starter-pack next. Otherwise it asks whether to schedule the autopilot, and at what cadence and time, always your choice, never assumed, printing one line like Autopilot scheduled: weekdays at 07:00 (via cowork).

Good to know

Section titled “Good to know”
  • If no AssureSwarm MCP server is detected, it tells you to add your company’s server at Claude.ai > Settings > Connectors, then re-run: it won’t write a config pointing at nothing.
  • It never invents your subdomain or server name: it detects what’s available and asks when there’s ambiguity, and it probes for reachability before writing, because a config aimed at an unreachable server is worse than no config.
  • Credentials are never written to config.json: authentication lives in the MCP connection at your Claude account level.
  • The autopilot schedule adapts to your runtime. Where the environment can create recurring tasks (Claude Cowork, Claude Code), setup registers the /coach-autopilot schedule for you; where it can’t (e.g. ChatGPT Codex), it prints exact instructions to add the scheduled task yourself. Either way the choice is recorded in the config, and re-running setup lets you keep, change, or remove it. A scheduling hiccup never fails setup.
  • Config is per-workspace and per-instance: keep dev, test, and prod tenants in separate working directories.
Section titled “Related”