Document back office
Document back office, planned call by call: what a coding agent asks Rhumb for this kind of app, and what the day's catalogs answer.
No chat. Receipts and invoices are read each night and written to QuickBooks or Xero, with bank data through Plaid and files from Google Drive and Gmail.
1. What the nightly job connects to.
POST /v1/stack
{
"systems": [
"google-drive",
"gmail",
"quickbooks",
"xero",
"plaid"
],
"target": "cloud",
"actions": [
"read",
"write"
]
}What must hold
- Every system named is known.
- Every system has a server, a way it is reached, or a stated "closed".
On the day's catalogs
- Google Drive
- server https://drivemcp.googleapis.com/mcp/v1
- Gmail
- server https://gmailmcp.googleapis.com/mcp/v1
- QuickBooks
- Its owner lists no server of its own in the MCP Registry.; reached: api
- Xero
- server https://api.xero.com/mcp/v1
- Plaid
- server https://api.dashboard.plaid.com/mcp/
- Warnings
- no_first_party (quickbooks), tools_not_read (google-drive), tools_not_read (gmail), tools_not_read (xero), tools_not_read (plaid)
- Skills
- none: no owner of these systems publishes one, and another publisher's is never named
2. Reading receipts.
GET /v1/resolve?purpose=read-receipts&providers=openai,anthropic,geminiWhat must hold
- A model is chosen.
- Without a stated volume, the cost is one call's.
On the day's catalogs
- Model
- gpt-5.6 on openai
- Typical call
- $0.012
- Reason
- cheapest vision.extract model within 5% of the best score (lmarena:document_style_control) on your providers; cheapest of its 3 offerings (openai); 14 cheaper models have no score to compare
- Warnings
- price_unverified, cheaper_unscored
3. The receipts model, served in the EU.
GET /v1/resolve?purpose=read-receipts&providers=openai,anthropic,gemini®ions=euWhat must hold
- An offering that states no region is warned of, not silently kept.
On the day's catalogs
- Model
- gpt-5.6 on openai
- Typical call
- $0.012
- Reason
- cheapest vision.extract model within 5% of the best score (lmarena:document_style_control) on your providers; cheapest of its 3 offerings (openai); 14 cheaper models have no score to compare
- Warnings
- price_unverified, region_not_filtered, cheaper_unscored
4. The receipts model again, with the month's volume stated.
GET /v1/resolve?purpose=read-receipts&providers=openai,anthropic,gemini&calls_per_user_month=3000What must hold
- A model is chosen.
- With the volume stated, a month's figure and its arithmetic.
On the day's catalogs
- Model
- gpt-5.6 on openai
- Typical call
- $0.012
- A user's month
- $36 (0.012 × 3,000 calls = 36)
- Reason
- cheapest vision.extract model within 5% of the best score (lmarena:document_style_control) on your providers; cheapest of its 3 offerings (openai); 14 cheaper models have no score to compare
- Warnings
- price_unverified, cheaper_unscored
5. Classifying transactions overnight, at the batch price.
GET /v1/resolve?purpose=classify-transactions&providers=openai,anthropic,gemini&service_tier=batchWhat must hold
- The price is the batch price, or the refusal says no batch price is held.
On the day's catalogs
- Model
- gemini-4-carbon on gemini
- Typical call
- $0.00105
- Reason
- cheapest text.chat model within 5% of the best score (lmarena:text_instruction_following) on your providers; 8 cheaper models have no score to compare
- Warnings
- cheaper_unscored
6. Where the nightly job runs.
POST /v1/platform
{
"requires": [
"scheduled_jobs"
]
}What must hold
- No major platform is warranted: the answer is a generic host.
- A free tier that sleeps under a scheduled job is warned of.
On the day's catalogs
- Kind
- any host meeting the need: 6 listed
- A host must
- A host must run jobs on a schedule.
- Warning
- Supabase Edge Functions and Render web services stop serving on a free tier left idle, and a scheduled job that wakes nothing will not run on it: use a paid plan for scheduled work.
See every use case, and Architect an agent to try a stack.