How to connect FreeScout to Anthropic Claude
Works with Claude
Connecting FreeScout to Claude takes one settings change in ReplyRabbit: choose Anthropic as a mailbox's provider, paste an Anthropic API key, pick a model, and save. From then on that mailbox's drafts, and on paid plans its next actions and AI Signals, are written by Claude. Anthropic is a paid-plan provider, so the mailbox needs ReplyRabbit Pro or Team.
Why teams pick Claude for support drafts
Claude models have a reputation for careful, natural-sounding replies, which is what you want when the draft is going to a customer with only a quick edit. In ReplyRabbit that shows up as drafts that stick to the question asked and don't over-promise. The right way to judge it is on your own tickets, and because providers are set per mailbox, you can run Claude on one inbox and compare it with another provider on a second one.
What you need
- ReplyRabbit Pro or Team on the mailbox. On the free plan the Anthropic option is greyed out.
- An Anthropic account with API access and an API key. Anthropic bills that account for usage; ReplyRabbit adds nothing on top.
- Company context and response style already filled in, so you're testing Claude with your business facts rather than a blank slate. See How to set up ReplyRabbit.
Connect Anthropic
Open ReplyRabbit for the mailbox
In FreeScout, open the mailbox, choose Settings, then ReplyRabbit.Choose Anthropic
In the AI Provider card, set Provider to Anthropic.Paste the API key
Paste your Anthropic API key into API Key. It's stored encrypted and shown masked after saving.Pick a Claude model
Use the latest option or a listed preset. Choose Custom only if your team has a specific model ID it has already tested.Save and generate a draft
Save, open a customer conversation, and generate a draft. Compare it with what your agent would have written.
Prompt caching happens on its own
Anthropic supports prompt caching, and ReplyRabbit uses it when it sees likely reuse across matching draft requests. Stable parts of the prompt, such as your company context and style rules, can be cached, while the customer's conversation text stays outside the cached prefix. If Anthropic rejects the cache control for a model, ReplyRabbit retries the plain prompt, so drafting fails open rather than stopping.
There's nothing to switch on. Note that the mailbox's Monthly Token Budget counts total input and output tokens even when Anthropic bills cached input at a lower rate, so the budget stays conservative.
Embeddings need another provider
Anthropic doesn't offer an embeddings endpoint, so if you use semantic search over your knowledge base or product catalogue, choose a different provider in the Embeddings subsection of AI Provider. Drafts stay on Claude; only retrieval uses the other key. Per-task AI models explains how the subsections share keys.
Troubleshooting
| Symptom | What to check |
|---|---|
| Anthropic can't be selected | The mailbox is on the free plan. Local and non-OpenAI providers need Pro or Team. |
| Invalid key | Re-copy the key from the Anthropic console and confirm it's from the account you expect. |
| Cache control error in logs | ReplyRabbit already retried without caching. Choose a supported Claude model if it repeats. |
| Semantic retrieval fell back to keywords | Set an embedding provider other than Anthropic. |
See Troubleshooting ReplyRabbit for the general cases, or compare providers in Choosing an AI provider.
Official references
Provider names, model lists, pricing, privacy terms, and legal guidance change. Check the current source before setting a production policy.