Orchard Discovery — Claude Code plugin

Point Claude Code at an MSP estate. Orchard watches the technicians' activity stream and the ConnectWise ticket/time corpus, and this plugin hands an agent the whole thing: the recurring manual work that shouldn't be manual — ranked, priced, and mapped to the connector steps that would automate it — plus the tickets, margins, fleet, signals and playbooks underneath it, and the rails to build and edit the automations themselves.

This is the customer-facing surface of Orchard's own MCP server: the same capability the platform runs on a schedule, callable by an agent, scoped to your tenant.

What you get

• /find-automatable-work — scans the estate and returns ranked, prescriptive automation candidates (with time and dollars saved per week).
• /discovery-report [estate] — scans, then renders a branded, deck-ready HTML Discovery Report artifact, via the bundled discovery-report skill.
• /automate [what] — finds the work and builds the automation: the agent authors a workflow graph against Orchard's node catalog + validator and saves it as a draft in your tenant (you activate it). Orchard supplies the rails; the agent does the authoring — no Orchard-side AI. Needs the build scope.
• /workflow-edit [playbook] — [change] — reads an existing playbook, proposes the change, shows you the node-level diff, validates, and saves a new version.
• /playbook-audit — every playbook checked for drift, fragility and dead weight: does the saved graph still validate, what keeps healing, what's active but never fires.
• /estate-brief [client] — a grounded brief on the estate: where the hours go, what needs attention, what's working, and the one thing to do next.
• /triage [severity] — open signals, failed runs and pending self-heals turned into a short ordered list of what a person should do now.
• /margin-review [client] — the money view: agreement margin, unlogged work, and what automation has measurably returned.

Two skills come along for the ride: discovery-report (the HTML artifact) and workflow-authoring (the graph grammar and the arming boundary, used by /automate and /workflow-edit).

• MCP tools, usable directly or by any agent. These are the server's own names; Claude Code namespaces tools per server, so installed from the marketplace they surface as mcp__plugin_orchard-discovery_orchard__<tool> (and as mcp__orchard__<tool> if you wire the server into .mcp.json by hand):
  • Orientation — estate_overview: fleet and client counts, workflows by status, open signals, which connectors are wired, how stale each source is, the governance posture, and what your connection is allowed to do. Start here.
  • Discovery — find_automatable_work (a fresh scan of activity + PSA), list_discoveries (the current candidates, no rescan).
  • Authoring & editing — get_workflow_toolkit (node catalog, wired connectors, authoring rules), validate_workflow (validator + linter + the diff of a proposed edit; no save), build_workflow (save a graph as a draft), update_workflow (edit an existing playbook), set_workflow_status (pause/stop/shelve one). The last three need the build scope.
  • Automation — list_workflows, get_workflow (the current graph), list_workflow_versions, get_workflow_evidence (the Witness's receipts), list_runs, get_run (the per-step trail), workflow_estate (pulse and drift), savings_ledger (measured savings vs. the frozen baseline), list_patterns, list_heals, list_documents (SOPs).
  • Fleet — list_machines, get_machine (with hardware inventory and live metrics), list_clients (or one client's impact rollup), activity_summary, list_activity, estate_graph (org / process / network), get_insights.
  • PSA — list_tickets, get_ticket (with its full timeline), psa_workload, agreement_margin, unlogged_work.
  • Governance — list_signals, governance_state, list_audit, list_team.

Setup — two lines, then sign in

In Claude Code:

/plugin marketplace add Orchard-HQ/orchard-mcp-plugin
/plugin install orchard-discovery@orchard

Then run any command (e.g. /find-automatable-work). The first call bounces you to Sign in with Orchard: your browser opens, you approve the connection in the Orchard dashboard while signed in as an owner/admin, and Claude Code receives a short-lived, tenant-scoped access token automatically. Nothing to configure, no token to paste. It refreshes itself; revoke access any time from the dashboard. To let the agent build automations, approve the build scope when prompted.

curl -X POST "https://api.entertheorchard.ai/v1/mcp-tokens" \
  -H "Authorization: Bearer <operator-jwt>" -H "Content-Type: application/json" \
  -d '{"name": "claude-code"}'   # returns "token": mcp_<tenanthex>.<secret>, shown once

Notes

• Tenant-scoped and safe. The token encodes your tenant; every tool is bound to your MSP's data by row-level security — never another tenant's, never the cross-tenant god-view.
• Writes are few, and scoped. find_automatable_work runs a scan, which upserts discovery candidates (an idempotent write). Only three other tools write anything — build_workflow, update_workflow, set_workflow_status — and all three require the build scope you approve at connect time. Everything else is read-only. Agent writes land in your own audit trail, actored as mcp-agent.
• The arming boundary: an agent authors, a human arms. This is enforced in the server, not asked for in a prompt:
  • build_workflow saves a draft. It cannot run until a person activates it.
  • update_workflow on a draft adopts the edit. On a live playbook it appends the new version but leaves the workflow pointed at the one a human approved — so an agent can stage a change to a running automation, never make one.
  • set_workflow_status can only lower autonomy (pause, disable, archive). An agent can pull the brake on something misbehaving; it cannot start anything.
• Accepting is an operator action. Turning a candidate into a live workflow happens in the Orchard console (Discoveries → Accept) under owner/admin sign-in — deliberately not an agent-executable tool, since it creates an automation that can act on real systems.
• Revoke a token any time: DELETE $ORCHARD_API_URL/v1/mcp-tokens/{id} (operator auth).

Development

There's no build step — this repo is markdown and JSON, and it installs straight from GitHub. python3 scripts/validate.py (stdlib only) checks the manifests, the command and skill frontmatter, and that no command names a tool the tool list above doesn't document. CI runs it on every push; Orchard's control plane runs the same script with the MCP server's real tool list, so a renamed tool fails there before it ships here.