Support agent
Support agent, planned call by call: what a coding agent asks Rhumb for this kind of app, and what the day's catalogs answer.
A hosted agent, run for many customers, that reads support tickets and writes replies. It works in Zendesk, Intercom, Slack, Notion, Shopify and Stripe, and searches the help documents.
1. What the agent connects to.
POST /v1/stack
{
"systems": [
"zendesk",
"intercom",
"slack",
"notion",
"shopify",
"stripe"
],
"target": "cloud",
"tenancy": "multi",
"actions": [
"read",
"write",
"untrusted_input"
]
}What must hold
- Every system named is known.
- Every system has a server, a way it is reached, or a stated "closed".
- Reading untrusted input with write access is warned of.
- Every server that takes one fixed key is warned of, since the agent runs for many customers.
- No skill is offered twice.
- The stack can be watched.
On the day's catalogs
- Zendesk
- Its owner lists no server of its own in the MCP Registry.; reached: api
- Intercom
- server https://mcp.intercom.com/mcp
- Slack
- server https://mcp.slack.com/mcp
- Notion
- server com.notion/mcp
- Shopify
- Its owner lists no server of its own in the MCP Registry.; reached: api
- Stripe
- server com.stripe/mcp
- Warnings
- untrusted_input_with_write, no_first_party (zendesk), no_first_party (shopify), tools_not_read (intercom), tools_not_read (slack), tools_not_read (notion), tools_not_read (stripe)
- Skills
- stripe-bfcm-dashboard (Stripe), upgrade-stripe (Stripe), stripe-best-practices (Stripe)
2. Embeddings for searching the help documents.
GET /v1/resolve?purpose=embed-search&providers=openai,anthropic,geminiWhat must hold
- An embedding model is chosen.
On the day's catalogs
- Model
- gemini-embedding-001 on gemini
- Typical call
- $0.000075
- Reason
- cheapest embedding model within 5% of the best score (mteb:rteb) on your providers; 1 cheaper model has no score to compare
- Warnings
- cheaper_unscored
3. Reranking search results, on providers that sell none.
GET /v1/resolve?purpose=rerank-search&providers=openai,anthropic,geminiWhat must hold
- The refusal names the nearest provider that sells one.
On the day's catalogs
Nothing meets the need: No candidate on these providers: providers left none. Nearest: nvidia/llama-nemotron-rerank-vl-1b-v2 on deepinfra, which fails providers.
4. The model that writes replies, made by a maker in the EU.
GET /v1/resolve?purpose=agent-with-tools&providers=openai,anthropic,gemini,mistral&origin=euWhat must hold
- Every model's maker is in the EU, or the refusal names the constraint.
On the day's catalogs
- Model
- mistral-large-2512 on mistral
- Typical call
- $0.0052
- Reason
- cheapest text.chat model within 5% of the best score (lmarena:text_style_control) on your providers from fr/de/it/es/nl/se/fi/ie/at/be/pl/pt/dk/cz/ee/lu makers; 10 cheaper models have no score to compare
- Warnings
- cheaper_unscored
See every use case, and Architect an agent to try a stack.