AI models in each client’s workflows
Four providers, each on the client’s own key: Anthropic’s Claude, OpenAI, Groq and OpenRouter. A workflow asks a model one question and passes the reply on, or hands an agent step a task and the few app steps it may use.
One prompt, one reply
Each provider has a step that sends one prompt and waits for the answer. The prompt is written on the step and can carry values from the run: the enquiry someone typed into the client’s form, a ticket’s subject, the notes from a booking. The model is set on the step, so two workflows for the same client can use different ones.
The reply comes back as the step’s output, ready for the next step by name: the text itself, why the model stopped (a reply cut short says so), the tokens it used and the model that answered. A later step can put the reply in an email, post it to Slack, write it into a CRM field or branch on it.
Each client connects a key, stored encrypted and bound to that client, so one client’s usage never lands on another’s bill and no template or pack carries a key with it.
What each provider takes
| Provider | Step | Model | Worth knowing |
|---|---|---|---|
| Anthropic | Ask Claude | Chosen from a list: Claude Opus 5.5 (the default) or Claude Sonnet 5.5, with an effort setting of low (fastest), medium or high (most thorough). | Up to 16,000 tokens of reply and two minutes per call. A request the model declines is run again on a fallback model inside the same call. |
| OpenAI | Ask a model | Any model id from the client’s OpenAI dashboard, typed in. | The key is made in the OpenAI project that bills this client, so the cost lands on the client’s account. |
| Groq | Ask a model | A model id from Groq’s model list, typed in. | Groq’s step is the quick one: it waits a minute at most, and since a question changes nothing, a call that times out is asked again. |
| OpenRouter | Ask a model | Any provider/model id listed on openrouter.ai/models, typed in. | One key reaches the models of many providers, and a credit limit set on that key caps what this client can spend. |
The agent step
An agent step gives the model a task and a short list of app steps it may call, six at most, each with the values you fix in advance (the Slack channel, the ClickUp list) and the few it may choose itself. It runs on Claude or on any of the three OpenAI-style providers, with the client’s key.
By default it stops after 5 model calls and 8 tool calls, which can be raised to 8 and 12 at most, and a run gets 5 minutes. Reaching a limit fails the step and says which, so a half-finished answer never moves on. A model error before any tool call is retried twice; after a tool has acted, nothing is retried.
Each tool call is a real request to the app, recorded in the run’s timeline as it happens, and the step ends with a line such as “Agent answered: 2 model calls, 1 tool call, 280 tokens”.
Library workflows with an AI step
Each of these playbooks includes an agent step, named here as the library names it, and installs into a client switched off with a checklist that asks for the model and the key.
- After-hours answering
Its AI step: “Answer from the facts”
Answer customer messages at any hour from the business’s own facts, and hand anything else to a person with a holding reply.
- Applicant intake
Its AI step: “Check the must-haves”
Acknowledge every job application at once, check it against the must-haves, log it, and send it to whoever hires.
- Booking brief
Its AI step: “Write the brief”
When someone books, send the owner a four-line brief: who they are, what they want, what they said, and one question worth asking.
- Conversation triage
Its AI step: “Sort the message”
Sort every new conversation in the support chat inbox by topic, page someone for the urgent ones, and put each in front of the team that handles it.
- Enquiry triage
Its AI step: “Sort the enquiry”
Sort every enquiry by what it is, so the owner’s phone only rings for the kinds that need the owner and the rest land with the office.
- Meeting follow-up
Its AI step: “Write the recap”
Turn the notes from a client meeting, pasted into a form, into a recap the client receives and the actions as a task.
- Request to task
Its AI step: “Write the task”
Turn every request posted in the requests channel into a task in the team’s list, and reply in the channel with the link.
- Ticket triage
Its AI step: “Sort the ticket”
Acknowledge every support ticket at once, sort it by topic and urgency, page someone for the urgent ones, and tell the team.
Packs that include one of them: Product-led software company, Sales-led software company, Executive coach, Small-business coach, Dropshipping store, Own-brand store, Subscription store, Barbershop stylist, Paid-ads agency, SEO agency, Retainer dev agency and Project-based dev agency.
Anthropic, Claude, OpenAI, Groq and OpenRouter and their logos are trademarks of their owners, used here only to say what Thrum works with; no endorsement is implied.
Thrum is cooking.
Get early access and you’ll be among the first in. Early members get a launch perk.