Prerequisites
- A running Breadbox instance with a reachable
/mcpendpoint (HTTPS for remote Openclaw gateways). - An Openclaw gateway installed and running.
- Breadbox registered as an MCP server in Openclaw — follow the Openclaw tab on the MCP setup page.
- A Breadbox API key. Start with read-only scope and graduate to full access only after observing a few correct agent cycles. See API keys.
Workflow 1 — The 12-hour triage loop
Pattern: Every 12 hours, the main agent checks how large theneeds-review backlog is. If it’s small enough to handle in one pass, it delegates a sub-agent to work the batch. If the queue is large, the sub-agent works in 30-transaction pages until it drains to zero.
This is the automated version of the Single Routine Reviewer pattern — same logic, running on a schedule instead of on demand.
Scheduling in Openclaw
Draft — verify with current Openclaw docs before publishing. Openclaw supports cron-like scheduling for recurring agent tasks. The exact configuration block for scheduling varies by Openclaw version; check https://docs.openclaw.ai for the current schedule config syntax. The example below reflects the general scheduling primitives documented as of early 2026.
sessions_spawn (Openclaw’s sub-agent primitive) to delegate the review batch to a child agent. The child agent inherits the Breadbox MCP connection from the parent workspace.
Main agent prompt (paste-ready)
Drop this into the main agent’s system instructions inside Openclaw:Sub-agent prompt (paste-ready)
This is the prompt for the child agent that does the actual review work:Workflow 2 — Attention-needed alerts
Pattern: After each review pass, the agent scans the deferred (needs-human) transactions. If any match certain criteria — large charge, unknown merchant, a charge that deviates from a known subscription amount — it sends a notification through Openclaw’s connected messaging channel.
This lets Breadbox reach you in Slack, Telegram, or wherever you’re already paying attention, instead of requiring you to check a dashboard.
Criteria for an alert
Common thresholds to configure (tune to your household):| Criterion | Example threshold |
|---|---|
| Large single charge | Amount > $200 |
| Unknown merchant | No prior transactions from merchant in last 90 days |
| Subscription deviation | Recurring merchant with amount +/- 15% of 3-month average |
| Duplicate candidate | Same merchant, same amount, same day as another transaction |
Notification channels
Openclaw exposes amessage() tool that routes outbound messages through any channel the gateway manages — Slack, Telegram, Discord, WhatsApp, iMessage (via BlueBubbles), and others. The call shape is consistent across channels:
"channel": "discord" for slack, telegram, etc., and set to to whatever channel/user identifier that platform uses (channel:<id>, user:<id>, thread:<id>). The agent prompt below uses that shape directly.
Channel routing — which channel ID belongs to “the household alerts channel” — is gateway-specific and configured where you set up your Openclaw tenant. See https://docs.openclaw.ai for the current per-channel reference if you need a different target format.
Alert agent prompt (paste-ready)
Run this agent after the triage loop completes (or as a standalone step inside the main agent before closing the session):Workflow 3 — Session hygiene
Pattern: Every write cycle starts withcreate_session and ends with submit_report. This is not optional overhead — it’s how Breadbox attributes agent work to specific cycles in the audit log. Without it, all agent writes are anonymous and the household has no clear record of what any given run did.
Why this matters
Breadbox’s Sessions reference explains that every write tool (other thancreate_session itself) requires a . Sessions group related tool calls into a unit visible in the admin dashboard. When Openclaw runs nightly and names its session “nightly triage — 2026-04-23”, the household can see exactly which transactions were touched, by which agent, and why.
submit_report closes the loop by posting a human-readable summary to the Breadbox Reports page. It is the agent’s equivalent of leaving a note — the household skims the title, reads the body if they want detail, and can see at a glance whether the last cycle ran cleanly.
Session open / close pattern
Every Openclaw agent that writes to Breadbox should follow this shell:Report format
Follow the style fromsubmit_report: the title is the primary signal — write it as a complete past-tense sentence with concrete numbers. The body uses ## headers, bullet lists, and inline transaction links for items that need a human look.
Example title: "Triage — 2026-04-23: 42 reviewed, 3 deferred, 1 anomaly flagged."
Cost, safety, and failure modes
Start read-only
Configure Openclaw’s Breadbox API key as read-only for the first few cycles. Watch the reports in the Breadbox dashboard — verify the agent is querying the right transactions, making sensible decisions, and correctly identifying what to defer. Promote to full access only after you trust the behavior. See Read-only vs. read-write.Guard against runaway writes
Even after a promotion to full access, a few defaults limit blast radius per cycle:- Keep
limit=30per page onquery_transactions. Smaller batches = smaller blast radius per cycle. - Review
submit_reportoutput for the first few weeks. If the deferred count is unexpectedly high or low, inspect the annotations before giving the agent more autonomy. - Breadbox’s
update_transactionsrespectscategory_override = true— rows a human has manually categorized are skipped automatically. Do not override this behavior.
Observe before automating
Use Breadbox’s Reports and Annotations pages as the primary observability surface:- Reports — every
submit_reportcall appears here, showing what the agent did each cycle. - Annotations — every tag change, category set, and comment is recorded on the individual transaction. Drill into any transaction to see the full agent reasoning trail.
Failure resilience
If Openclaw loses its machine, crashes, or loses its API token between cycles, the only harm is missed cycles. The data Breadbox holds is safe — no in-flight transaction updates are lost becauseupdate_transactions is atomic per call. The next time Openclaw recovers and runs, it will find the same backlog waiting.
If a cycle errors mid-way, the needs-review tags on unprocessed transactions remain intact. The household can see the half-finished session in the Breadbox admin dashboard and investigate from there.
Related reading
- Set up MCP for Claude and AI agents — configure the Breadbox MCP connection in Openclaw.
- MCP reference overview — full list of available tools and scopes.
- Single Routine Reviewer — the manual version of the triage loop this guide automates.
- Multi-agent reviewer flows — if the queue grows too large for a single agent pass.