erpipe-org / odoo-agency-fleet-review

Odoo agency fleet review

You are answering questions across a fleet of client Odoo databases through
one odoo-mcp server with named instances. Every result is tagged with its
_instance; one client being down must never sink the whole answer.

Prerequisites

• Multi-instance config (ODOO_CONFIG_FILE with an instances map).
  list_instances shows names, tags, and which allow cross-instance reads
  ("cross_instance": false opts a client out — respect it silently).

• Cross-instance tools are read-only by design. Writes happen per-instance
  through the normal gate, one client at a time.

Playbook

1. Map the fleet: list_instances — report count, tags, default, and
   any opted-out clients (just the count, not a complaint).

2. Fleet health: accounting_health_across_instances(instances="all")
   (or {"tags": ["managed"]}). For fleets >10, run via
   submit_async_task and poll.

3. Triage the errors map first. The response carries per-instance
   errors — an unreachable client is a finding in itself (report it,
   with diagnose_odoo_call output if the human wants the cause), not a
   reason to retry the whole fan-out.

4. Rank and drill down. Present a per-client table sorted by the metric
   the human asked about (e.g. overdue AR). For the worst clients, drill
   down with instance-scoped calls:
   receivable_payable_aging(instance="client_x"),
   search_records(..., instance="client_x").

5. Cross-client comparisons stay honest: averages are deliberately not
   merged across instances (different currencies/configs) — compare counts
   and per-client aggregates, never invent a fleet-wide average.

6. Per-client actions: anything beyond reading switches to that single
   instance and goes through the write gate there. Approval tokens are
   instance-bound; never reuse one across clients.

Output format

Fleet summary (reachable/unreachable/opted-out counts), the ranked
per-client table with _instance labels, drill-down findings, and a
follow-up list grouped by client.

Hard rules

• Never name an opted-out instance's data in results — it is opted out.

• Each instance runs under its own field ACL and rate budget; if a client's
  response says redacted_fields, that is policy, not an error.

• Fan-out is bounded (ODOO_MCP_CROSS_INSTANCE_WORKERS); for very large
  fleets prefer the async path over repeated synchronous sweeps.